Vol. 115 No. 8
Article
Administering Section 2 of the Voting Rights Act After Shelby County

Administering Section 2 of the Voting Rights Act After Shelby County

Christopher S. Elmendorf* & Douglas M. Spencer**

Until the Supreme Court put an end to it in Shelby County v. Holder, section 5 of the Voting Rights Act was widely regarded as an ef­fective, low-cost tool for blocking potentially discriminatory changes to election laws and administrative practices. The provision the Supreme Court left standing, section 2, is generally seen as expensive, cumber­some, and almost wholly ineffective at blocking changes before they take ef­fect. This Article argues that the courts, in partnership with the Department of Justice, could reform section 2 so that it fills much of the gap left by the Supreme Court’s evisceration of section 5. The proposed reformation of section 2 rests on two insights: first, that national survey data often contains as much or more information than precinct-level vote margins about the core factual matters in section 2 cases; and sec­ond, that the courts have authority to regularize section 2 adjudication by creating rebuttable presumptions. Most section 2 cases currently turn on costly, case-specific estimates of voter preferences generated from precinct-level vote totals and demographic information. Judicial deci­sions provide little guidance about how future cases—each relying on data from a different set of elections—are likely to be resolved. By creat­ing ev­identiary presumptions whose application in any given case would be determined using national survey data and a common sta­tistical model, the courts could greatly reduce the cost and uncer­tainty of section 2 litigation. This approach would also reduce the dependence of vote dilution claims on often-unreliable techniques of ecological infer­ence and would make coalitional claims brought jointly by two or more minority groups much easier to litigate.

* Professor of Law, University of California, Davis.

** Associate Professor of Law and Public Policy, University of Connecticut. For helpful discussions and feedback the authors thank Eric Biber, Guy Charles, Jack Chin, Heather Gerken, Mike Gilbert, Perry Grossman, Rick Hasen, Ellen Katz, Michael McDonald, Jennifer Nou, Rick Pildes, Mike Pitts, Kevin Quinn, David Schleicher, Nick Stephanopoulos, Franita Tolson, and discussants and participants at the 2014 Conference on Empirical Legal Studies, the 2014 Annual Meeting of the Midwest Political Science Association, and the Political Economy and Public Law Conference at the University of Rochester. We are also indebted to Lynn Vavreck and Simon Jackman for sharing the 2008 Cooperative Campaign Analysis Project dataset. Replication materials and appendix are available at http://www.dougspencer.org/research.html.

INTRODUCTION

  1. SECTION 2 AS A WEAK SUBSTITUTE FOR SECTION 5
    1. Conventional Wisdom About Sections 2 and 5
    2. Looming Threats to Section 2
  2. MAKING IT WORK: PRESUMPTIONS FOR THE CORE OF SECTION 2
    1. Preliminaries: Interpreting Section 2 in Constitutional Context (After Shelby County)
    2. Presumptions About “Risk Factors” for Constitutional Violations
      1. Racial Attitudes and Beliefs
      2. Racial Polarization in Partisanship
    3. Presumptions About Disparate Impact
      1. Vote Dilution
        1. Preference / Interest Divergence (“Polarization”)
        2. Minority Opportunity Districts
      2. Vote Denial
      3. Summary
  3. AUTHORITY, LEGAL AND OTHERWISE
    1. Legal Authority to Create the Presumptions
    2. Judicial Competence
  4. MODEL BUILDING AND RESULTS
    1. Tools for Estimating Racial-Group Opinion Within Subnational Geographic Units
    2. An Illustrative Model and Maps
    3. Next Steps

CONCLUSION

Introduction

Widely lauded as one of the most effective statutes ever enacted, the Voting Rights Act of 1965 (VRA) finally made good on the promise of the Fifteenth Amendment. 1 See, e.g., Charles S. Bullock III & Ronald Keith Gaddie, The Triumph of Voting Rights in the South 323 (2009) (“The consensus is that the [VRA] has been inordinately successful.”); Samuel Issacharoff et al., The Law of Democracy: Legal Structure of the Political Process 514–16 (4th ed. 2012) (noting “guarantees of the Fourteenth and Fifteenth Amendments were essentially disregarded in many states” before passage of VRA, which “transformed American politics in a variety of ways”). The VRA outlaws the use of “tests or devices” as a prerequisite to voting, and section 2 of the statute further prohibits state and local governments from structuring elections “in a manner which results” in members of a group defined by race or color “hav[ing] less opportunity than other members of the electorate to participate in the political process and to elect representatives of their choice.” 2 52 U.S.C. § 10301 (Supp. II 2015). The VRA was formerly codified in various sec­tions of Title 42, but all of its provisions have been moved to the new Title 52 as part of an editorial reclassification. U.S. House of Representatives, Office of the Law Revision Counsel, Editorial Reclassification Title 52, U.S. Code, http://uscode.house.gov/‌editorialreclassification/t52/index.html [http://perma.cc/‌W9DQ-HMBJ] (last visited Oct. 14, 2015). Sections 4 and 5 target states and localities with a history of black dis­en­franchisement, requiring these states to obtain prior approval from the federal government before implementing any changes to their elec­tion laws. The principal question in these “preclearance” proceedings is a simple one: Would the change make minority voters worse off? 3 Preclearance could also be denied if the change was adopted for discriminatory rea­sons. See 52 U.S.C. § 10304(a) (mandating target jurisdictions seek preclearance and demonstrate proposed restrictions “neither [have] the purpose nor will have the effect of” abridging voting rights on basis of race (emphasis added)). The ju­risdiction seeking preclearance bears the burden of proving it would not.

In June 2013, the Supreme Court in Shelby County v. Holder put the preclearance mechanism on ice. 4 133 S. Ct. 2612, 2631 (2013). The Court found the coverage formula (which determines the states and localities subject to preclearance) fa­cially unconstitutional, faulting Congress for not updating the formula when Congress reauthorized section 5 in 2006. 5 Id. at 2628–29, 2631. Justice Kennedy mused that section 5 was probably not needed in any event because discrimina­tory voting changes can also be blocked, preimplementation, by prelim­inary injunctions in lawsuits brought under section 2. 6 Transcript of Oral Argument at 37, Shelby County, 133 S. Ct. 2612 (No. 12-96). Leading election lawyers disagree. 7 See, e.g., Rick Hasen, “Voting Law Decision Could Sharply Limit Scrutiny of Rules,” Election Law Blog (Feb. 28, 2013, 8:00 am), http://electionlawblog.org/?p=47852 [http://perma.cc/58SW-UMJN] (“Justice Kennedy seems to mistakenly believe that section 2 liability plus preliminary injunctions would be just as good as section 5 liability.”); J. Gerald Hebert & Armand Derfner, More Observations on Shelby County, Alabama, and the Supreme Court, Campaign Legal Ctr. Blog (Mar. 1, 2013), http://www.campaign‌legalcenter.org/news/blog/more-observations-shelby-county-alabama-and-supreme-court [http://perma.cc/9MBU-V27L] (arguing Court overestimated efficacy of preliminary in­junctions under section 2). Section 2 litigation is costly and rarely results in pre­liminary relief; 8 Hebert & Derfner, supra note 7 (“The actual number of preliminary injunctions that have been granted in the hundreds of Section 2 cases that have been filed over the years is quite small, likely putting the percentage at less than 5%, and possibly quite lower.”). moreover, Shelby County further undermines the already shaky constitutional moorings of section 2. 9 See infra section II.A (describing constitutional requirements for section 2 claims); cf. Richard Hasen, The Curious Disappearance of Boerne and the Future Jurisprudence of Voting Rights and Race, SCOTUSblog (June 25, 2013, 7:10 pm), http://scotusblog.com/2013/06/the-curious-disappearance-of-boerne-and-the-future-jurisprudence-of-voting-rights-and-race/ [http://perma.cc/A42B-9R5T] (“[T]he Court for now seems to have foreclosed greater deference for voting decisions under Congress’s Fifteenth amendment powers[, which] could spell trouble for Section 2 of the Voting Rights Act . . . and other laws aimed at preventing race discrimination in voting.”).

Shelby County’s impact was felt immediately. A number of states that had been subject to the preclearance process quickly adopted or imple­mented new, restrictive voting laws. For example, the day Shelby County was decided, Texas announced that it was implementing its strict voter-ID (identification) requirement, which section 5 had previously blocked. 10 Ryan J. Reilly, Harsh Texas Voter ID Law “Immediately” Takes Effect After Voting Rights Act Ruling, Huffington Post (June 25, 2013, 2:04 pm), http://www.huffingtonpost.
com/2013/06/25/texas-voter-id-law_n_3497724.html [http://perma.cc/‌437C-RCJW]. Texas also announced that its legislature’s 2011 redistricting maps would immediately take effect. Id. Preclearance had been denied because a three-judge panel of the District Court for the District of Columbia was “persuaded by the totality of the evidence that the plan was en­acted with discriminatory intent.” Texas v. United States, 887 F. Supp. 2d 133, 161 (D.D.C. 2012). Voter-ID laws recently adopted in Alabama 11 Ala. Code § 17-9-30 (2014). and Virginia 12 Va. Code Ann. § 24.2-643 (2015). were also freed to take effect. 13 Alabama had passed the photo-ID law in 2011 but had not requested preclear­ance because the Secretary of State’s office had yet to develop rules for implementing the law. Kim Chandler, State Has Yet to Seek Preclearance of Photo Voter ID Law Approved in 2011, AL.com (June 12, 2013, 7:30 am), http://blog.al.com/wire/2013/06/photo_voter_id.html [http://perma.‌cc/FC6P-ELRH]. Two months later, North Carolina enacted a sweep­ing election reform bill that the president of the state’s National Association for the Advancement of Colored People (NAACP) chapter called, “the worst voter suppression law since the days of Jim Crow.” 14 Zachary Roth, N.C. Wants Voting Law Emails Kept Secret, MSNBC (Jan. 9, 2014, 2:54 pm), http://www.msnbc.com/msnbc/nc-wants-voting-law-emails-kept-secret [http://‌perma.cc/4VNY-2GUK]. North Carolina’s Voter Information Verification Act takes effect January 1, 2016. N.C. Gen. Stat. § 163-166.13 (Supp. 2014). Dur­ing the same month, Mississippi passed new ID requirements for vot­ing. 15 Miss. Code Ann. § 23-15-563 (2015). Local governments freed from preclearance also made some im­portant changes. The city of Pasadena, Texas, replaced two district coun­cil seats in predominately Latino neighborhoods with two at-large seats elected from the majority-white city. 16 SCOTUSblog, After Shelby County, YouTube, at 00:25 (Nov. 4, 2013), https://www.youtube.com/watch?v=u_Eo97odboQ [http://perma.cc/AN25-CD2D]. Galveston County, Texas, cut in half the number of constable and justice-of-the-peace districts, eliminating virtually all of the seats currently held by Latino and black incumbents. 17 John Suayan, Lawsuit: New Galveston Co. Electoral Map ‘Racially Discriminatory,’ Se. Tex. Rec. (Aug. 27, 2013, 11:49 am), http://setexasrecord.com/‌stories/510622002-lawsuit-new-galveston-co-electoral-map-racially-discriminatory [http://‌perma.cc/H874-FQ7C]. And the city of Macon, Georgia, moved the date of city elections from November to July, when black turnout has traditionally been low. 18 Ga. Voters Surprised Macon Election Change Isn’t Challenged, NPR Morning Edition (Feb. 6, 2014, 5:00 am), http://www.npr.org/2014/02/06/272359791/voting-right
s-act-update (on file with the Columbia Law Review).

With Congress divided and slow to respond to Shelby County, 19 It was not until January 2014 that the civil rights community and its allies in Congress came forth with draft legislation responding to Shelby County. Voting Rights Amendment Act of 2014, H.R. 3899, 113th Cong. § 2 (2014). The bill would bring only four of the formerly covered states back under section 5, see Summary of the Voting Rights Amendment Act of 2014, Advancement Project, http://www.advancementproject.‌org/pages/summary-of-the-voting-rights-amendment-act-of-2014-introduced-january-16-20#sthash.fPoBAqJb.dpuf [http://perma.cc/8PX3-UC7J] (last visited Sept. 2, 2015) (noting proposed coverage formula would likely cover only Texas, Mississippi, Georgia, and Louisiana), not­withstanding that much broader coverage (using a different formula) is readily justified, see Christopher S. Elmendorf & Douglas M. Spencer, The Geography of Racial Stereotyping: Evidence and Implications for VRA Preclearance After Shelby County, 102 Calif. L. Rev. 1123, 1166–69 (2014) [hereinafter Elmendorf & Spencer, Preclearance] (showing coverage formula based on racial stereotyping could reach most of the formerly covered jurisdictions). then-Attorney General Holder pledged to do all he could to protect voting rights using the remnants of the VRA. 20 See Sari Horwitz, Justice Department to Challenge States’ Voting Laws, Wash. Post (July 25, 2013), http://www.washingtonpost.com/politics/justice-department-to-chall
enge-states-voting-rights-laws/2013/07/25/c26740b2-f49b-11e2-a2f1-a7acf9bd5d3a_story.html [http://perma.cc/Y2TM-MSB2] (“‘[W]e plan . . . to fully utilize the [VRA’s] remaining sections to ensure the voting rights of all American citizens are protected.’”). The Department of Justice (DOJ) challenged new voting restrictions in Texas and North Carolina under section 2, 21 Josh Gerstein, DOJ Challenges N.C. Voter ID Law, Politico (Sept. 30, 2013, 12:03 am), http://www.politico.com/story/2013/09/justice-department-north-carolina-voter-id-law-97542.html [http://perma.cc/2PVQ-2F3Z]; Press Release, Dep’t of Justice, Justice Department to File New Lawsuit Against State of Texas over Voter I.D. Law (Aug. 22, 2013), http://www.justice.gov/opa/pr/justice-department-file-new-lawsuit-against-state-texas-over-voter-id-law [http://perma.cc/Q5G5-9J37]. and other private and public lawsuits are in the offing. 22 See Horwitz, supra note 20, at 2 (noting DOJ’s plan to utilize sections 2 and 3 of VRA). For a summary of election law changes in formerly covered jurisdictions since Shelby County, see States and Localities’ Responses to Shelby County, Alabama v. Holder, http://www.naacpldf.org/document/states-responses-shelby-decision [http://perma.cc/V7AT-PZ
NZ] (last visited Oct. 14, 2015).

This Article takes up the question of whether section 2 can be made to function like erstwhile section 5 in the post–Shelby County world. We argue that it can—provided that courts, litigators, and DOJ come to un­derstand two fundamental points. First, national survey data often con­tain as much or more information about core evidentiary matters in section 2 cases than the precinct-level vote tallies and demographics that have been the grist of voting rights litigation for the last generation. De­mographic and legal changes are undermining the conventional sources of evidence for many section 2 cases, but at the same time, advances in survey administration, reweighting, and model-based estimation of local political preferences from national surveys are generating new kinds of evidentiary materials that speak to the central factual questions in these cases. Second, because section 2 is a “common law statute” (or statutory provision), the courts have authority to create rebuttable presumptions to guide and regularize the adjudication of section 2 claims. 23 Regarding section 2 as a common law statute, see Christopher S. Elmendorf, Making Sense of Section 2: Of Biased Votes, Unconstitutional Elections, and Common Law Statutes, 160 U. Pa. L. Rev. 377, 448–55 (2012) [hereinafter Elmendorf, Making Sense of Section 2].

We show that the courts could create rebuttable presumptions un­der section 2 that would give the statute special bite in many jurisdictions for­merly covered by section 5. Implemented with national survey data rather than local election tallies, the new presumptions would greatly reduce the cost and uncertainty of challenging under section 2 the kinds of elec­tion law changes that DOJ used to block under section 5. The presump­tions would also go a long distance toward establishing the “likelihood of success on the merits” needed for preliminary relief. 24 See Montana v. Suffolk Cty. Legislature, 268 F. Supp. 2d 243, 260 (E.D.N.Y. 2003) (applying likelihood-of-success standard in vote dilution case).

Even if the courts decline our call to create formal evidentiary pre­sumptions under section 2, mere judicial recognition of the fact that na­tional survey data shed light on the central factual questions in section 2 cases would breathe new life into the statute. Presently, section 2 cases are like snowflakes. Judicial rulings on the evidentiary materials in one case provide little direction for the next case down the pike because these ma­terials vary so much from case to case (often the courts emphasize voting patterns in local elections). But if the same national data sets were de­ployed in case after case, the ordinary processes of common law adjudi­cation would generate substantial guidance about whether any given would-be defendant is likely to be held liable under section 2. This is so whether or not the courts create de jure evidentiary presumptions.

*     *     *

Plaintiffs in a section 2 case must establish, first, that the challenged election law, procedure, or practice has a racially disparate impact. 25 Infra text accompanying note 98. Sec­ond, plaintiffs must connect this impact to “social and historical condi­tions.” 26 Infra text accompanying note 98. In view of Shelby County, the required “social and historical con­ditions” showing is best understood as a way of testing whether the rem­edy the plaintiff seeks is proportionate to the present-day risk of uncon­stitutional race discrimination in the electoral process. Thus, we shall generally call it the “constitutional risk” requirement, rather than using the more conventional “social and historical conditions” label.

The presumptions we propose address both prongs of a section 2 claim: constitutional risk and disparate impact. The constitutional-risk requirement should be deemed rebuttably satisfied if the jurisdiction’s majority-group citizens subscribe to substantially negative stereotypes of the minority, or if there is an extremely high correlation between race and reliably partisan voting in the defendant jurisdiction (provided that actors affiliated with the white-preferred party were responsible for the election law or practice at issue). 27 See infra sections II.B.1–II.B.2. By shifting the burden of persuasion to defendants, the courts acknowledge that partisan motives do not merit the same presumption of legitimacy in jurisdictions where the partisan payoff to racial discrimination is exceptional. 28 Where there is an extreme correlation between race and partisanship, opposing-party actors have incentives to target and burden voters on the basis of their race, race being easier to observe than reliable partisanship. See generally Adam B. Cox & Richard T. Holden, Reconsidering Racial and Partisan Gerrymandering, 78 U. Chi. L. Rev. 553, 572–76 (2011) (demonstrating black voters’ liberal ideological distribution leads both Democrats and Republicans to treat them differently than white voters in redistricting).

As for the disparate-impact prong, one must distinguish so-called “vote dilution” cases, which address the rules for aggregating votes into representation, from “vote denial” cases, which concern barriers to cast­ing a valid, duly counted ballot. 29 See Daniel P. Tokaji, The New Vote Denial: Where Election Reform Meets the Voting Rights Act, 57 S.C. L. Rev. 689, 702–03 (2006) (distinguishing “first generation” VRA litigation focused on vote denial claims from “second generation” litigation attacking “practices that diminish minorities’ voting strength where they were permitted to vote”). In dilution cases, districting schemes and at-large electoral systems are generally said to have a disparate im­pact if—given white and minority political preferences—they prevent the minority community from electing a “roughly proportional” number of minority “candidates of choice” and additional, reasonably compact majority-minority districts could be drawn. 30 See J. Gerald Hebert, Paul M. Smith, Martina E. Vandenberg & Michael B. DeSanctis, The Realist’s Guide to Redistricting: Avoiding the Legal Pitfalls 33–61 (2d ed. 2010) [hereinafter Hebert et al., Realist’s Guide] (“Although rough proportionality does not automatically protect a state from liability under Section 2, it is a strong ‘indication that minority voters have an equal opportunity, in spite of racial polarization, to partici­pate in the political process and to elect representatives of their choice.’” (internal quota­tion marks omitted) (quoting Johnson v. De Grandy, 512 U.S. 997, 1020 (1994))); Ellen Katz et al., Documenting Discrimination in Voting: Judicial Findings Under Section 2 of the Voting Rights Act Since 1982, 39 U. Mich. J.L. Reform 643, 730–32 (2006) (discussing proportionality and noting in nearly all cases in which lower courts made finding on pro­portionality, liability question was resolved accordingly); cf. League of United Latin Am. Citizens v. Perry, 548 U.S. 399, 436 (2006) (beginning “totality of circumstances” inquiry by assessing proportionality, implying it is especially important factor to consider). Racial minorities have “can­didates of choice” if and only if the minority is internally politically cohe­sive. 31 See Hebert et al., Realist’s Guide, supra note 30, at 48–50 (stating minority politi­cal cohesiveness can be demonstrated by “‘showing that a significant number of minority group members usually vote for the same candidates’” (quoting Thornburg v. Gingles, 478 U.S. 30, 56 (1986) (plurality opinion) (Brennan, J.))). And unaccommodating electoral designs threaten the minority’s opportunity to elect only if political preferences are racially polarized, meaning that a politically cohesive racial majority opposes the minority’s preferred candidates. 32 See id. at 45–48 (describing presence of racially polarized political preferences and voting as “linchpin” of vote dilution cases because “[i]f there is no racially polarized voting, ‘it cannot be said that the ability of minority voters to elect their chosen represent­atives is [less than] that of white voters’” (alteration in original) (quoting Gingles, 478 U.S. at 48 n.15 (plurality opinion) (Brennan, J.))).

Presumptions for the disparate-impact side of a vote dilution claim must therefore address racial polarization in political preferences and the related question of whether a given electoral district is a “minority opportunity district” (MOD)—one that gives minority voters the oppor­tunity to elect their “candidates of choice.” 33 See League of United Latin Am. Citizens [LULAC], 548 U.S. at 427–36 (applying Gingles framework and using term “opportunity district” to describe districts in which minority voters have realistic chance to elect candidates of their choice); Hebert et al., Realist’s Guide, supra note 30, at 34–37 (introducing “Gingles test” for vote dilution). Though polarization has un­til now been gauged on the basis of voting patterns in local elections, this Article contends that courts may rebuttably presume racial polarization if plaintiff-race minority citizens in the defendant jurisdiction substantially diverge from other citizens in terms of their policy preferences, general political ideology, or socioeconomic status. 34 Defendants could rebut the polarization inference with data from local elec­tions, but it would not be necessary for plaintiffs to introduce local voting data after estab­lishing presumptive polarization. See infra section II.C. We also offer a presumptive definition of “minority opportunity district,” based on demographics.

Presumptions about disparate impact are harder to craft for vote de­nial cases, in part because this body of law is still in its infancy. This Article offers two very tentative suggestions, focusing on the correlation between race or ethnicity and socioeconomic status, and on demograph­ic divergence between the populations of eligible and actual voters.

Unlike the coverage formula for section 5 preclearance, there would be no de jure list of jurisdictions “covered” by the section 2 presump­tions. Rather, plaintiffs would have to make evidentiary showings at the start of their case to establish which presumptions apply. We demonstrate in Part IV that most of these showings could be made using multilevel statistical modeling and data from existing national surveys, such as the National Annenberg Election Survey, the Cooperative Congressional Election Study, and the Cooperative Campaign Analysis Project. The em­pirical results in Part IV suggest that blacks (but not necessarily other racial groups) are likely to be protected by the presumptions throughout the Deep South, that is, in most of the formerly covered jurisdictions. The two fastest growing racial groups in the United States, Asian Americans and Latinos, are jointly politically cohesive almost everywhere. This implies that Asians and Latinos ought to have considerable success bringing “coalitional” vote dilution claims under section 2—which has not been the case to date. 35 See generally Ming Hsu Chen & Taeku Lee, Reimagining Democratic Inclusion: Asian Americans and the Voting Rights Act, 3 U.C. Irvine L. Rev. 359, 390–91 (2013) (find­ing vote dilution suits brought by Asian Americans have been “overwhelmingly un­success­ful,” and “the handful of claims involving Asian Americans all . . . concern . . . the applica­tion of VRA to multi-racial and ‘other minority’ groups”); Chelsea J. Hopkins, Comment, The Minority Coalition’s Burden of Proof Under Section 2 of the Voting Rights Act, 52 Santa Clara L. Rev. 623, 625 (2012) (noting in coalitional claims “courts have ex­pressed resistance to granting relief without a heightened showing of group cohesion”). Note that the continued availability of coalitional claims under section 2 is open to ques­tion after Bartlett v. Strickland, 556 U.S. 1 (2009) (plurality opinion) (Kennedy, J.). See infra note 284. However, our results also indicate that Asian American and Latino plaintiffs may find it harder than blacks to satisfy the “constitutional risk” requirement.

Although the approach suggested in this Article would not yield an of­ficial list of jurisdictions covered by the presumptions, a pattern of de facto coverage should emerge as courts and litigants come to a shared understanding of what the presumptions are and how they may be estab­lished in a given case. As models and data sources become standardized (more on this below), litigants will be able to see which jurisdictions face presumptive liability.

State and local officials in the de facto covered jurisdictions would have to disprove central elements of a section 2 case, much as covered jurisdictions bore the burden of proof in preclearance proceedings un­der section 5. This shifting of evidentiary burdens should make it fairly easy for plaintiffs to obtain preimplementation preliminary relief, much as DOJ under section 5 was able to block suspicious changes before they took effect. In a redistricting case, for example, plaintiffs could establish the requisite “likelihood of success” by showing that the defendant failed to create a roughly proportional number of presumptive opportunity dis­tricts when it was feasible to do so. Or, in a challenge to voter-ID re­quire­ments, plaintiffs might obtain preliminary relief by showing that the law was adopted on a substantially party-line vote (thereby establishing par­tisan intent), that the burden of the law would fall mostly on low-income voters, and that race and poverty are highly correlated. Because defend­ants would have to rebut the inference of discrimination where the rele­vant presumptions apply, section 2 litigation would be costlier for defend­ants than for plaintiffs, incentivizing defendants to settle quickly and on terms favorable to the plaintiffs. Lawmakers and election administrators in these jurisdictions would have correspondingly strong ex-ante incen­tives to safeguard minority voting rights. 36 The recently introduced Voting Rights Amendment Act of 2014 also aims to fa­cilitate preliminary relief under section 2, but in a different manner. Though the amend­ments are not entirely clear on this point, it appears they would replace the traditional four-prong test for a preliminary injunction with a simple weighing of relative hardship (to defendants and to plaintiffs), without any consideration of the plaintiffs’ likelihood of suc­cess on the merits. See Voting Rights Amendment Act of 2014, H.R. 3899, 113th Cong. § 6(b)(4) (2014). Whether this is constitutional is an open question. At best, it permits plaintiffs to maintain the status quo while a suit proceeds, without the information-forcing and settlement-inducing benefits of our proposal.

The balance of this Article unfolds as follows. Part I provides a brief overview of sections 2 and 5. It also explains the conventional wisdom that (weak, cumbersome) section 2 is no substitute for (potent, efficient) section 5, as well as the less widely appreciated fact that section 2 may not be able to play in the future even the limited role it has played in the past due to recent developments in law and in statistics.

Parts II, III, and IV develop this Article’s proposal for a presumption-driven section 2. Part II identifies what we take to be the central factual questions in section 2 cases and explains how they could be answered using evidentiary presumptions and survey data. Part III steps back and considers the courts’ authority to create the presumptions suggested in Part II. We argue that judicial authority to establish the presumptions is pretty straightforward as a matter of law. However, judges may under­stand­ably be reluctant to exercise this authority without technical guid­ance from an expert agency. DOJ, assisted by an outside advisory panel, could provide a useful nudge, issuing non-binding guidelines about data sources, statistical techniques, and appropriate legal inferences. Part IV turns to empirical methods and results. We introduce the art and science of multilevel regression with poststratification (MRP), a recently devel­oped tool for estimating local opinion from national survey data, and we present some initial results and maps, highlighting regions of the country likely to be covered de facto by the presumptions. The online appendix provides further information about MRP. 37 Christopher S. Elmendorf & Douglas M. Spencer, Technical Appendix, Administering Section 2 of the VRA After Shelby County, http://www.dougspencer.org/research/s2_Appendix.pdf [http://perma.cc/C5Y5-N4MU] [hereinafter Elmendorf & Spencer, Technical Appendix] (last updated Jan. 7, 2015).

The Supreme Court has long interpreted the VRA pragmatically. 38 See generally Guy-Uriel E. Charles & Luis Fuentes-Rohwer, The Voting Rights Act in Winter: The Death of a Superstatute, 100 Iowa L. Rev. 1389, 1410–20 (2015) (de­scribing weak foundations in text and legislative history of many canonical opinions). That pragmatism is now feared by many to be the Act’s undoing, 39 See id. at 1420–38 (predicting further unraveling of Act in absence of normative consensus about its aims). but it is also grist for the reconstruction offered here. The demise of the VRA is not inevitable.

I. Section 2 as a Weak Substitute for Section 5

To frame this Article’s proposal, we begin by outlining the standard understandings of sections 2 and 5, the conventional wisdom that section 2 is weak and ineffective in comparison to section 5, and the looming threats to section 2 as it has been implemented to date.

A. Conventional Wisdom About Sections 2 and 5

The potency of section 5 is commonly attributed to its substitution of administrative for judicial procedures; its establishment of a fairly bright-line results test; and, critically, its placement of the burden of proof on the party seeking preclearance. 40 For an excellent summary of the procedural and substantive differences between sections 2 and 5 with more detail than provided here, see generally Nicholas O. Stephanopoulos, The South After Shelby County, 2013 Sup. Ct. Rev. 55, 62–86 [hereinafter Stephanopoulos, The South After Shelby County]. Congress’s delegation of authority to DOJ to make preclearance decisions meant that the determinations could be made with a minimum of legal expenses, for covered jurisdic­tions and would-be plaintiffs alike. 41 See id. at 65 (“[T]he DOJ shoulders much of the expense of preclearance under Section 5.”). Though covered jurisdictions were permitted to opt out of DOJ review in favor of a preclearance proceeding before the District Court for the District of Columbia, 52 U.S.C. § 10304(a) (Supp. II 2015), this option was rarely invoked, as it was much more costly for the jurisdiction seeking preclearance. See Stephanopoulos, The South After Shelby County, supra note 40, at 65 (discussing expense of judicial preclearance in contrast to inexpensive administrative preclearance).

The principal substantive standard under section 5 was reasonably clear cut. Preclearance was to be denied if the measure “‘would lead to a retrogression in the position of racial minorities with respect to their ef­fective exercise of the electoral franchise’” 42 Georgia v. Ashcroft, 539 U.S. 461, 466 (2003) (quoting Beer v. United States, 425 U.S. 130, 141 (1976)). or if it was adopted with a discriminatory purpose. 43 52 U.S.C. § 10304(b)–(c). Discriminatory intent can be hard to prove (or disprove), 44 As a leading treatise notes, “the criteria used to evaluate a plan under Section 5’s purpose prong [were] vast and comprehensive.” Hebert et al., Realist’s Guide, supra note 30, at 29. Also, the “intent” standard under section 5 was essentially dormant from the date of the Supreme Court’s decision in Reno v. Bossier Parish School Board until the enact­ment of the 2006 Amendments to the VRA, which expressly abrogated that decision. Compare Reno v. Bossier Parish Sch. Bd., 528 U.S. 320, 341 (2000) (holding section 5 “does not prohibit preclearance of a redistricting plan enacted with a discriminatory but non-retrogressive purpose”), with Fannie Lou Hamer, Rosa Parks, and Coretta Scott King Voting Rights Act Reauthorization and Amendments Act of 2006, Pub. L. No. 109-246, 120 Stat. 577, 578 (codified at 52 U.S.C. §§ 10302–10305, 10308–10310, 10503) (criticizing Reno and stating Reauthorization Act abrogates that decision). but the “retrogressive effects” prong of section 5 did a lot of the work. Congress boiled the retrogression inquiry down to the question of whether the electoral change would hinder minorities’ ability to elect their “preferred candidates of choice.” 45 Voting Rights Act Reauthorization and Amendments Act, 120 Stat. at 580–81. DOJ and the courts denied pre­clearance when a change would reduce the number or reliability of elec­toral districts that provide minorities with an opportunity to elect minor­ity candidates 46 See, e.g., Texas v. United States, 887 F. Supp. 2d 133, 149–51 (D.D.C. 2012) (in­ter­preting section 5 to protect only those districts that give politically cohesive minority com­munity opportunity to elect their preferred candidates, as opposed to simply oppor­tunity to elect Democrats). or would create a material barrier to voting borne dispro­portionately by minority citizens. 47 See, e.g., Texas v. Holder, 888 F. Supp. 2d 113, 144 (D.D.C. 2012) (denying pre­clearance to Texas voter-ID requirement because law would create significant barrier to voting for poor people and racial minorities were disproportionately represented among poor); Florida v. United States, 885 F. Supp. 2d 299, 329 (D.D.C. 2012) (denying preclear­ance to Florida’s reduction in early voting days because African Americans disproportion­ately used early voting and reduction represented “material burden” on franchise); Letter from Thomas E. Perez, Assistant Att’y Gen., U.S. Dep’t of Justice, to Keith Ingram, Dir. of Elections, Office of the Tex. Sec’y of State 5 (Mar. 12, 2012), http://www.brennancenter.org/sites/default/files/analysis/DOJ_Final_Letter_To_Texas_On_Voter_ID_Law.pdf [http://perma.cc/X8GY-ZMVE] (denying preclearance because state’s own data showed Hispanics disproportionately lacked qualifying ID and statute did not adequately mitigate burdens on voters lacking qualifying ID). According to law professor and for­mer DOJ staff attorney Michael Pitts, “local officials and their demogra­phers” in the covered jurisdictions were “acutely cognizant of the stand­ards for approval and typically tr[ied] to steer very clear of anything that would raise concerns with the Attorney General.” 48 Michael J. Pitts, Let’s Not Call the Whole Thing Off Just Yet: A Response to Samuel Issacharoff’s Suggestion to Scuttle Section 5 of the Voting Rights Act, 84 Neb. L. Rev. 605, 613–14 (2005) [hereinafter Pitts, Let’s Not Call the Whole Thing Off].

Finally, because section 5 put the burden of proof on the party seek­ing preclearance, the provision was information-forcing. 49 This point is emphasized in Guy-Uriel E. Charles & Luis Fuentes-Rohwer, Mapping a Post-Shelby County Contingency Strategy, 123 Yale L.J. Online 131, 137 (2013), http://yalelawjournal.org/pdf/1172_7tf1ew4q.pdf [http://perma.cc/T7KM-U7CH] [hereinafter Charles & Fuentes-Rohwer, Post-Shelby Contingency Strategy] (arguing preclearance “facili­tates monitoring through disclosure” by “forc[ing] the institutions with the best information about potential discriminatory practices to share that information”). Jurisdictions contemplating an election law change that might disadvantage racial mi­norities had incentives to gather information about potentially retrogres­sive impacts and to mitigate those impacts ex ante. 50 For one such case study, see Ellen D. Katz, South Carolina’s “Evolutionary Process,” 113 Colum. L. Rev. Sidebar 55, 58–61 (2013), http://www.columbialawreview.‌org/wp-con­tent/uploads/2013/03/Katz-113-Colum.-L-Rev.-Sidebar-55.pdf [http://perma.cc/4F4G-WQYY] (concluding grant of preclearance of South Carolina’s voter-ID law resulted from “evo­lutionary process” of state officials addressing concerns regarding law’s impact on mi­nority voters). If DOJ remained worried about potential impacts, it could respond with a “More Information Request,” essentially putting the new law on hold until the state or local government had gathered enough information to allay DOJ’s concerns. 51 Regarding the enforcement function of More Information Requests, see Luis Ricardo Fraga & Maria Lizet Ocampo, More Information Requests and the Deterrent Effect of Section 5 of the Voting Rights Act, in Voting Rights Act Reauthorization of 2006: Perspectives on Democracy, Participation, and Power 47, 52–53, 65–67 (Anna Henderson ed., 2007) (discussing procedural steps for More Information Requests and impact on compliance).

The world of section 5, then, was a world in which civil rights advo­cates could block voting changes that might disadvantage the minority community without spending huge sums of money on courtroom legal fees, expert witnesses, and the like. For advocacy groups worried about a change in local election procedures, it was often enough to fire off a let­ter outlining their concerns to DOJ. DOJ lacked the resources to give in-depth scrutiny to each of thousands of preclearance proceedings, so it relied on community groups to flag changes that merited special scru­tiny. 52 See Heather K. Gerken, A Third Way for the Voting Rights Act: Section 5 and the Opt-In Approach, 106 Colum. L. Rev. 708, 725–26 (2006) (explaining DOJ initiates re­sponse to preclearance requests by placing “informal call . . . to a civil rights group or an elected minority official to see if there is a problem” with proposed restriction). Some attorneys general were probably more solicitous of minority communities than others, 53 Some commentators have worried that partisan political considerations play an ex­cessive role in preclearance determinations, particularly for congressional and statewide redistrictings. See, e.g., Samuel Issacharoff, Is Section 5 of the Voting Rights Act a Victim of Its Own Success?, 104 Colum. L. Rev. 1710, 1730–31 (2004) (suggesting use of preclear­ance for districting may permit partisan gain unlikely to be noticed or addressed). For thoughtful replies to this line of critique, see Luis Fuentes-Rohwer & Guy-Uriel E. Charles, The Politics of Preclearance, 12 Mich. J. Race & L. 513, 534–35 (2007) (arguing “ebbs and flows of politics” are not “as noxious a development as some commentators believe” in voting rights context); Ellen D. Katz, Democrats at DOJ: Why Partisan Use of the Voting Rights Act Might Not Be So Bad After All, 23 Stan. L. & Pol’y Rev. 415, 419–23 (2012) (sug­gesting partisan use of VRA generates viable claims and responds to existing racial dis­putes); Pitts, Let’s Not Call the Whole Thing Off, supra note 48, at 630 (“Partisan in­trigue . . . does not constitute a compelling enough reason for section 5 to slip softly out of existence.”). but to the extent that DOJ cared about mi­nority voting rights, the structure of section 5 made the path from “be­coming concerned” to “blocking the change” easy and inexpensive to navigate.

The contrast with section 2 could not be more dramatic: Section 2 dis­putes are adjudicated in judicial rather than administrative fora, the legal standard for liability under section 2 is murky, and the burden of proof falls on the party challenging the election law at issue rather than the party defending it. 54 See Stephanopoulos, The South After Shelby County, supra note 40, at 62–66 (sum­marizing procedural differences between sections 2 and 5); id. at 74–86 (discussing sub­stantive differences in vote-dilution context); id. at 106–10 (discussing substantive dif­fer­ences in vote-denial context).

Substantively, section 2 prohibits electoral arrangements “which re­sult[]” in members of a class of citizens defined by race or color “hav[ing] less opportunity than other members of the electorate to par­ticipate in the political process and to elect representatives of their choice.” 55 52 U.S.C. § 10301 (Supp. II 2015). The courts have struggled to flesh out this abstraction. It is gen­erally agreed that plaintiffs must show (1) that the election law at is­sue has a racially disparate impact and (2) that this impact can be chalked up to the law’s “interaction” with “social and historical condi­tions.” 56 See infra section II.A (discussing requirements to show proof of section 2 violations). Most courts and commentators also agree that section 2 sup­ports independent causes of action for “vote denial” and “vote dilution.” 57 As a reminder, denial cases concern barriers to the casting of valid, duly counted ballots; dilution cases concern rules for aggregating votes into representation, such as districted versus at-large elections. See supra text accompanying notes 29–32.

But these generalities conceal much normative uncertainty and dis­a­greement. The ultimate question in section 2 cases is whether the “total­ity of circumstances” warrant a finding of liability. 58 Johnson v. De Grandy, 512 U.S. 997, 1017–18 (1994). These “circum­stances” include (but are not limited to) the defendant jurisdiction’s his­tory of discrimination, lingering effects of past de jure discrimination, racial appeals in political campaigns, racially polarized voting, informal barriers to ballot access for minority candidates, unusual features of the electoral system that may disadvantage minorities, and strength or weak­ness of the state interests asserted in defense of the challenged election laws. 59 These factors were enumerated in the Senate Judiciary Committee report accom­pa­nying passage of the section 2 results test. S. Rep. No. 97-417, at 28–29 (1982).

Still unresolved is the normative question to be answered when ex­am­ining the totality of circumstances. 60 See generally Christopher S. Elmendorf, Kevin M. Quinn & Marisa A. Abrajano, Racially Polarized Voting, 83 U. Chi. L. Rev. (forthcoming 2016) (manuscript at 36–50), http://papers.ssrn.com/‌sol3/papers.cfm?abstract_id=2668889 (on file with the Columbia Law Review) [hereinafter Elmendorf et al., Racially Polarized Voting] (discussing four theories of vote dilution and their conflicting implications). In dilution cases, some courts focus on whether the minority community can elect a “roughly propor­tional” number of its candidates of choice. 61 This factor was prioritized—without being made decisive—by the Supreme Court in De Grandy, 512 U.S. at 1013–14 & n.11 (“‘Proportionality’ as the term is used here links the number of majority-minority voting districts to minority members’ share of the rele­vant population.”). See generally Katz et al., supra note 30, at 730–32 (documenting lower court reliance on this factor). Other courts use the totality of circumstances inquiry to assess whether the plaintiffs’ injury can fairly be traced to disparate-treatment or intentional race discrimination, wheth­er by conventional state actors or nominally private actors. 62 See, e.g., United States v. Charleston County, 365 F.3d 341, 349 (4th Cir. 2004) (“[T]he reason for polarized voting is a critical factor in the totality analysis . . . .”); Goosby v. Town of Hempstead, 180 F.3d 476, 502–03 (2d Cir. 1998) (Leval, J., concurring) (“[W]here the complaint essentially alleges that voters of the protected class have had little success electing candidates of their choice, and where correction would require radical political restructuring . . . . I believe courts should not find a violation in the absence of race-based intent.”); Lewis v. Alamance County, 99 F.3d 600, 616 n.12 (4th Cir. 1996) (“We think the best reading of the several opinions in Gingles . . . is one that treats causation as irrelevant in the inquiry into the . . . preconditions, but relevant in the totality of circumstances in­quiry . . . .” (citations omitted)); S. Christian Leadership Conference of Ala. v. Sessions, 56 F.3d 1281, 1293–94 (11th Cir. 1995) (“[T]here was ample evidence in the record to sup­port the [district] court’s conclusion that factors other than [candidate] race, such as par­ty politics and availability of qualified [black] candidates, were driving the election re­sults . . . .”); Vecinos De Barrio Uno v. City of Holyoke, 72 F.3d 973, 982–83 (1st Cir. 1995) (explaining plaintiffs’ showing of strong correlation between voter race and vote choice gives rise to inference of racial bias on part of white electorate, which may be rebutted with evidence “voting patterns can most logically be explained by factors unconnected to the intersection of race with the electoral system”); Nipper v. Smith, 39 F.3d 1494, 1523–25 (11th Cir. 1994) (en banc) (plurality opinion) (Tjoflat, C.J.) (adopting similar presumption and burden-shifting rule); see also League of United Latin Am. Citizens, Council No. 4434 v. Clements, 989 F.2d 831, 849–63 (5th Cir. 1993) (en banc) (holding white bloc voting within meaning of third prong of Gingles is not “legally significant” unless “caused” by race). (This has become known as the section 2 causation requirement. 63 See D. James Greiner, Causal Inference in Civil Rights Litigation, 122 Harv. L. Rev. 533, 590–97 (2008) [hereinafter Greiner, Causal Inference] (discussing courts’ use of causation element when adjudicating section 2 cases); Katz et al., supra note 30, at 670–72 (discussing case law). As Greiner and Katz observe, some courts roll the “causation issue” into the Gingles polarized-voting inquiry, and others consider it part of the “totality of cir­cumstances.” Katz suggests—we think correctly—that little turns on this distinction. Id. at 671. ) And still other courts churn through the motions of the totality of circumstances analy­sis without stopping to explain their underlying conception of equal po­litical opportunity. 64 See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 37) (“Most judicial opinions in vote dilution cases are nominally atheoretical.”). In vote denial cases, the role of the totality of circum­stances inquiry is, if anything, even more unsettled. 65 Cf. League of Women Voters of N.C. v. North Carolina, 769 F.3d 224, 239 (4th Cir. 2014) (“[T]here is a paucity of appellate case law evaluating the merits of Section 2 claims in the vote-denial context.”). Compare Tokaji, supra note 29, at 709–26 (arguing most of Senate Report factors are not relevant to vote denial cases), with Veasey v. Perry, 29 F. Supp. 3d 896, 918–19 (S.D. Tex. 2014) (relying on Senate Report factors in case about voter-ID requirement). The factors high­lighted by Congress in enacting the results test were gleaned from earlier vote dilution cases, and courts and commentators disagree about how (if at all) they bear on vote denial cases. 66 See supra note 65. Problems of vote denial simply were not part of the congressional debates. 67 See Thomas M. Boyd & Stephan J. Markman, The 1982 Amendments to the Voting Rights Act: A Legislative History, 40 Wash. & Lee L. Rev. 1347, 1356–425 (1983) (re­counting legislative history, which was entirely focused on vote dilution).

What is clear is that section 2’s uncertain substantive norm, coupled with its express call for a totality of circumstances inquiry, has made liti­gating section 2 cases expensive and unpredictable. Plaintiffs must as­sem­ble local election data and hire statisticians to estimate voting pat­terns. 68 For local government elections, these data are rarely available in convenient elec­tronic formats and assembling the data is often a significant cost in section 2 litigation. See infra text accompanying note 157. Historians may be called to speak to past practices in the locale. Candi­dates, elected officials, and community leaders are asked to testify about their personal experiences with bloc voting, racial campaign ap­peals, and the like. 69 On the importance of qualitative evidence for vote dilution litigation under section 2, see Hebert et al., Realist’s Guide, supra note 30, at 48 (“Anecdotal evidence is often used [in Section 2 cases] to supplement statistical findings.”); D. James Greiner, Re-Solidifying Racial Bloc Voting: Empirics and Legal Doctrine in the Melting Pot, 86 Ind. L.J. 447, 484–87 (2011) [hereinafter Greiner, Re-Solidifying Racial Bloc Voting] (discuss­ing types of qualitative evidence that can help judges make inferences about racial polarization). The causation inquiry further complicates matters. Plaintiffs challenging a felon disenfranchisement rule, for example, may have to prove that the state’s penal code is administered in an intentionally dis­criminatory fashion. 70 See, e.g., Farrakhan v. Gregoire, 623 F.3d 990, 992 (9th Cir. 2010) (rejecting chal­lenge to Washington State’s felony disenfranchisement rule on ground that plaintiff failed to establish intentional discrimination in criminal justice administration). Worse yet, as Jim Greiner has explained, the causa­tion question has often been posed in ways that may render it unanswerable. 71 Greiner, Causal Inference, supra note 63, at 591–97.

Together, the fact-intensive nature of section 2 claims and the un­cer­tain standard for liability make preliminary relief hard to obtain. Vet­eran litigators estimate that plaintiffs have secured preliminary injunc­tions in only about 5% of section 2 cases. 72 Hebert & Derfner, supra note 7. The path from “becoming concerned” to “blocking a change” is slow and arduous. Meanwhile, officials elected under racially discriminatory ground rules may pass new laws that further hinder minority candidates or otherwise disadvantage the minority community.

To say that section 2 pales in comparison to section 5 is not to say that it is toothless. There has emerged a nascent ecosystem of civil rights groups that monitor state and local governments and have some in-house capacity for litigation. 73 See Charles & Fuentes-Rohwer, Post-Shelby Contingency Strategy, supra note 49, at 123–24 (describing litigation efforts of civic groups under section 5). Also, well-funded actors such as political parties and unions sometimes finance section 2 cases when the political stakes are high, for example, when the litigation could shift the balance of pow­er in a state legislature or in Congress. 74 Samuel Issacharoff has argued that political parties and associated actors typi­cally have the most to gain in section 2 cases. See Samuel Issacharoff, Gerrymandering and Political Cartels, 116 Harv. L. Rev. 593, 646 (2002) (describing “clear partisan gain” for political parties in redistricting litigation). However, the section 2 results test is under threat from two directions—one jurisprudential, the other demographic and statistical.

B. Looming Threats to Section 2

The Supreme Court has issued a string of decisions narrowing sec­tion 2 on the basis of the constitutional avoidance canon. 75 See Elmendorf, Making Sense of Section 2, supra note 23, at 399–404 (sum­mariz­ing cases). Shelby County provides an accelerant, as the Court’s rejection of Jim Crow history as the rationale for section 5 coverage casts doubt on the common judicial prac­tice of grounding section 2 “social and historical conditions” find­ings on the same history. 76 See infra section II.A (describing threshold for section 2 relief after Shelby County). More generally, the normative uncertainty at the heart of section 2 makes it difficult to assess whether the results test represents a congruent and proportional response to constitutional violations. 77 Elmendorf, Making Sense of Section 2, supra note 23, at 409–14 (evaluating and contrasting competing normative understandings of section 2).

The other rising threat to section 2 is that the statistical techniques used to establish minority political cohesion and white bloc voting tend to break down if there are more than two racial groups or significant res­idential integration in the jurisdiction—which is increasingly typical. Mi­nority cohesion and white-bloc voting have traditionally been inferred from aggregate rather than individual-level data (precinct-level election returns plus racial demographics from the Census). 78 For a great introduction to the statistical techniques used in vote dilution cases, see D. James Greiner, Ecological Inference in Voting Rights Act Disputes: Where Are We Now, and Where Do We Want to Be?, 47 Jurimetrics 115, 123–50 (2007) [hereinafter Greiner, Ecological Inference]. This works reasona­bly well when there are only two racial groups and precincts are racially homogenous. But as the number of racial groups increases from two to three or four, and as neighborhoods become less homogeneous, the amount of information about racial voting patterns in the precinct-level data becomes very sparse. 79 See D. James Greiner & Kevin M. Quinn, Exit Polling and Racial Bloc Voting: Combining Individual-Level and R×C Ecological Data, 4 Annals Applied Stats. 1774, 1776 (2010) [hereinafter Greiner & Quinn, Exit Polling and Racial Bloc Voting] (describing challenges to modeling voting patterns as diversity increases); Greiner, Re-Solidifying Racial Bloc Voting, supra note 69, at 465–68 (“[M]ore racial groups mean more moving parts.”). Conclusions about racial polarization under these conditions are tenuous—unless the analyst can supplement the ag­gregate data with individual-level observations obtained from exit polls and other surveys. 80 See Adam Glynn & Jon Wakefield, Ecological Inference in the Social Sciences, 7 Stat. Method. 307, 307 (2010) (demonstrating “inclusion of a small amount of individual level data can dramatically improve the properties of [the] estimates”); Greiner & Quinn, Exit Polling and Racial Bloc Voting, supra note 79, at 1777 (“[O]ur hybrid estimator al­lows inferences unavailable from either the exit poll or the ecological inference model alone.”). But survey data about vote choice in local elections “are almost never available” in vote dilution cases. 81 Bernard Grofman, Expert Witness Testimony and the Evolution of Voting Rights Case Law, in Controversies in Minority Voting: The Voting Rights Act in Perspective 197, 217 (Bernard Grofman & Chandler Davidson eds., 1992).

Courts may well start to reject section 2 claims on the ground that the evidence of racially polarized voting is unreliable. Would-be plaintiffs who suspect a section 2 violation may have to wait several election cycles before bringing suit, pouring money into exit polls all the while. 82 Greiner treats this as an unavoidable consequence of his results. See Greiner, Re-Solidifying Racial Bloc Voting, supra note 69, at 482 (“The need for polls over several elec­tion cycles may be a fact of life in some multiracial polities.”).

II. Making It Work: Presumptions for the Core of Section 2

Having set up the problem, we now elaborate our solution. The ar­gu­ment proceeds in three steps. The first step, which this Part develops, is to explain how central factual questions on which section 2 liability turns could, in principle, be translated into rebuttable evidentiary pre­sumptions, implemented cheaply and predictably using national survey data and standard statistical models. 83 National surveys rarely ask about vote choice in local elections. And even if the sur­vey did ask about local elections, the new statistical tools for estimating local opinion from national surveys (explained in Part IV) could not be used to estimate vote choice in local elections. The tools assume that all survey respondents have answered the same question. After establishing that the core of section 2 can be translated into such presumptions, we will address judi­cial authority to create the presumptions (Part III) and statistical tools for estimating local opinion using national surveys (Part IV).

Where the presumptions apply, defendants would have to rebut the in­ference of a statutory violation, much like state and local governments in preclearance proceedings under section 5 carried the burden of proof. Section 2 would therefore become information forcing in much the same way as section 5. And, assuming that the presumptions apply predictably and at low cost, it will be easy for civil rights groups and po­tential defendants to figure out ex ante who is likely to bear the burden of proof (and on which questions) in a section 2 case. 84 Though not quite so easy as looking up the jurisdiction’s name on the list of cov­ered jurisdictions under section 5. Potential defend­ants that are “covered” de facto by the presumptions would know this, and, like jurisdictions that were covered de jure under the preclearance regime, would have incentives to anticipate and mitigate potential dis­parate impacts whenever they change their election laws. 85 In one important respect, a presumption-driven section 2 would actually be more powerful than erstwhile section 5. Under section 5, state and local governments were never liable for leaving in place existing laws with a racially disparate impact; only changes to the jurisdiction’s electoral arrangements could be challenged on an “impact” (rather than intent) theory. See 52 U.S.C. § 10304(a) (Supp. II 2015) (applying when covered jurisdiction “shall enact or seek to administer” new voting restriction). By contrast, the results test of section 2 allows status-quo arrangements to be challenged. 52 U.S.C. § 10301(b). If they do not, plaintiffs should be able to obtain preliminary relief, even preimplemen­tation, as the presumptions would go a long distance toward establishing the necessary “likelihood of success on the merits.”

The use of national survey data and standard statistical models is very important for low-cost, predictable implementation. Instead of gath­ering case-specific datasets, prospective litigants should be able to go online, download the relevant dataset and statistical model, and figure out which presumptions apply. Because the same datasets and models would be used in case after case, each judicial decision would provide consid­erable guidance about the next case. This is not true of section 2 today because the cases tend to be litigated on the basis of voting patterns in local elections. 86 “Polarized voting” is the central issue in vote dilution cases, see Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 8–15) (reviewing evolution of ra­cial polarization “test”), and one factor among many that courts may consider in vote de­nial cases, see Christopher Elmendorf, Judge Easterbrook on the Voting Rights Act: Asking Good Questions, Making Bad Law, Election Law at Moritz (Oct. 8, 2014), http://moritzlaw.osu.edu/election-law/article/?article=12965 [http://perma.cc/P8YG-VM3Z] (noting racially polarized voting is important in vote denial cases due to considera­tion of social and historical conditions). Evidence of racial polarization or its absence in so-called “en­dog­enous” elections—elections to the governmental body at issue in the case—typically re­ceives the most weight. Elmendorf et al., Racially Polarized Voting, supra note 60 (manu­script at 25). The candidates and issues in these elections vary from one jurisdiction to the next, so a court’s determination in one case about, for example, what constitutes a “legally significant” level of polarized vot­ing generally does not carry over to the next case. 87 For a detailed examination of how the racially polarized voting inquiry works in practice today, see Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 15–35).

This argument comes with an important caveat. The benefits of im­plementing section 2 with rebuttable presumptions and national survey data—lower cost and more predictable litigation, with preliminary relief becoming easier to secure in parts of the country where the likelihood of racial discrimination with respect to voting is high—depend on wide­spread agreement about what the presumptions are and how they apply in a given case. But widespread agreement may be hard to achieve. The courts are not of one mind about the meaning of section 2, 88 See generally Elmendorf, Making Sense of Section 2, supra note 23, at 386–403 (de­scribing Supreme Court’s “ongoing uncertainty about the meaning of Section 2, the Court’s worries about the effects of Section 2 on racial politics, and the Court’s doubts about the constitutionality of Section 2”); Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 36–49) (summarizing federal judges’ “competing normative theories” in vote dilution cases). and even conditional on a particular gloss on section 2’s meaning, the presump­tions could be defined in a variety of ways. 89 See infra sections II.B–II.C (suggesting evidentiary presumptions). In addition, estimating local opinion from national survey data depends on various discretionary mod­eling choices, inviting “battles of the experts” even if the courts have reached agreement on how to define the presumptions. 90 See infra note 221 and accompanying text (discussing evidentiary showings for craft­ing rebuttable presumptions for section 2).

Guidance from an administrative agency such as DOJ may be neces­sary to solve the judicial coordination problem. 91 See infra notes 222–231 and accompanying text (suggesting possibility of greater DOJ action to limit judicial discretion in setting presumptions). For now, it is enough to show that important factual questions on which section 2 liability de­pends can, in principle, be answered using rebuttable presumptions and national survey data.

The argument of this Part unfolds as follows. Section II.A presents a Shelby County–informed gloss on the central elements of a section 2 claim. It argues that plaintiffs must show (1) that the challenged practice has a ra­cially disparate impact (the “disparate impact prong”) and (2) that the rem­edy sought is a reasonable response to recent, un­constitutional race dis­crim­ination or the present and future risk of un­constitutional race dis­crim­ination in the defendant jurisdiction (the “constitutional risk prong”). Section II.B suggests that the constitutional risk prong can be satisfied with evidence of racial stereotyping in the de­fendant jurisdiction or possibly with evidence of an extreme correlation between race and reliable partisan vot­ing. Section II.C sketches out some tentative presumptions for the disparate-impact prong, focusing on elec­toral district demographics (for dilution cases), and racial gaps in voter participation rates as well as impacts by socioeco­nomic class (for denial cases).

A. Preliminaries: Interpreting Section 2 in Constitutional Context (After Shelby County)

In order to translate the core factual questions under section 2 into ev­identiary presumptions, one must first establish what those questions are and whether there are specific legal constraints that the presump­tions must respect. These questions are tricky. As noted above, the statu­tory text is opaque and the evolving case law has not created much nor­mative clarity. 92 See generally Elmendorf, Making Sense of Section 2, supra note 23, at 387–95 (“There is certainly no plain meaning to this statutory standard, and the legislative history is opaque in critical respects.”).

Ours is a practical project, so rather than begin with some idealized account of section 2, we begin with an account that we think a broad range of judges could accept—including those in the center of the cur­rent Supreme Court. The Court has repeatedly signaled its discomfort with section 2, often using the constitutional avoidance canon to justify narrow constructions. 93 Id. at 379–403. The constitutional problem in a nutshell is this: Section 2 establishes a results test, but the Fourteenth and Fifteenth Amendments prohibit only subjective discrimination 94 A terminological note: The term “subjective discrimination” is used here as a catchall for intentional and disparate-treatment discrimination. Though the terms “inten­tional discrimination” and “disparate treatment discrimination” are sometimes used inter­changeably by the courts, disparate treatment—treating person A differently than person B because of A’s race—need not be intentional. It may result from biases or preferences of which the decisionmaker is unaware. See generally Symposium on Behavioral Realism, 94 Calif. L. Rev. 945 (2006) (covering various forms of implicit bias in human behavior and intersection with law). The Supreme Court’s recent decision in Texas Department of Housing and Community Affairs v. Inclusive Communities Project, Inc., 135 S. Ct. 2507 (2015), strongly signals that even unintended disparate treatment on the basis of race generally violates the Fourteenth Amendment. See Samuel R. Bagenstos, Disparate Impact and the Role of Classification and Motivation in Equal Protection Law After Inclusive Communities, 101 Cornell L. Rev. (forthcoming) (manuscript at 17) (on file with the Columbia Law Review) (noting Court “referred not just to ‘disguised animus’ but also to ‘un­conscious prejudices’” in describing discriminatory intent uncovered by disparate im­pact doctrine (quoting Inclusive Cmtys., 135 S. Ct. at 2522)). on the basis of race. 95 See Elmendorf, Making Sense of Section 2, supra note 23, at 399–403 (describing how constitutional concerns facilitate narrow reading of section 2 by conservative Justices). Though the Supreme Court has stated that Congress, when en­forcing the Fourteenth Amendment, may prohibit some conduct that does not violate the Constitution, the Court has imposed a tailoring re­quirement: Enforcement measures must be “congruent and propor­tional” to the record or threat of constitutional violations. 96 City of Boerne v. Flores, 521 U.S. 507, 520 (1997). It is doubtful that a pure disparate-impact standard could satisfy this requirement. 97 See Elmendorf, Making Sense of Section 2, supra note 23, at 400–01 (claiming ra­cially disproportionate election outcomes alone may not justify application of section 2 as “congruent and proportional” under Boerne).

But plaintiffs in a section 2 case must show more than disparate im­pact. Rather, as many courts have stated, plaintiffs must establish not only (1) that the challenged election law, procedure, or practice has a racially disparate impact on the minority’s opportunity to participate in the polit­ical process (in vote denial cases) or to elect representatives of its choice (in vote dilution cases) but also (2) that this disparate impact can be chalked up to the “interact[ion]” of the challenged provisions with “so­cial and historical conditions.” 98 Thornburg v. Gingles, 478 U.S. 30, 47 (1986) (plurality opinion) (Brennan, J.); see also League of Women Voters of N.C. v. North Carolina, 769 F.3d 224, 239 (4th Cir. 2014) (extending this idea to vote denial claims); Ohio State Conference of NAACP v. Husted, 768 F.3d 524, 554–55 (6th Cir. 2014) (same).

The answer to those who doubt section 2’s constitutionality lies in the murky “social and historical conditions” requirement. It should be interpreted to maintain a nexus between liability under the statutory re­sults test and actual or threatened constitutional violations. The showing ought to establish that unconstitutional race discrimination in the elec­toral process is at least “significantly likely” to occur or to have occurred in the defendant jurisdiction and that the remedy the plaintiff seeks is a reasonable response to the risk of constitutional violations. 99 See Elmendorf, Making Sense of Section 2, supra note 23, at 417–47 (stating plain­tiffs in section 2 cases must show significant likelihood of link between injury and “race-biased decisionmaking”). To avoid confusion, we suggest relabeling the “social and historical conditions re­quirement” as the “constitutional risk requirement.”

In an important sense, the “social and historical conditions” or “con­stitutional risk” showing is section 2’s counterpart to the coverage for­mula for section 5. Because the coverage formula for section 5 tracked historical, unconstitutional race discrimination, 100 Technically, the coverage formula used proxies: low rates of voter participation and the use of “tests or devices” to disqualify voters. But the formula was reverse engi­neered to cover most of the Jim Crow South. See South Carolina v. Katzenbach, 383 U.S. 301, 329–30 (1966) (describing coverage formula and its purpose). preclearance could be denied on the basis of retrogressive impact alone, without any further show­ing to establish a nexus between the remedy—denial of preclear­ance—and constitutional violations (at least until the coverage formula began to seem dated). But because section 2 applies nationally and thus to many state and local governments with no history of entrenched, de jure race discrimination, it is necessary for plaintiffs to make a further, case-specific showing that judicial intervention to remedy the racial dis­parity is justified.

A few caveats are in order. There is no settled judicial understanding of the “social and historical conditions” requirement in section 2 cases. Some judges have said it simply corroborates that the racially disparate im­pact “is not merely a product of chance.” 101 E.g., Frank v. Walker, 17 F. Supp. 3d 837, 876–77 (E.D. Wis. 2014), rev’d on other grounds, 768 F.3d 744 (7th Cir. 2014). This also seems to have been the understanding of the plurality in Gingles, 478 U.S. 30, which tried to boil vote dilution law down to the sim­ple question of whether minority-preferred candidates were almost always defeated. See id. at 46–51 (plurality opinion) (Brennan, J.). So understood, the “social and historical conditions” showing need not say anything about the nex­us between the remedy the plaintiffs seek and the risk of constitu­tional violations. Indeed, in most cases the showing would be redundant with the evidence of disparate impact. The impacts with which section 2 is con­cerned (minority representation and minority voter participation) result from the “interaction” of the challenged measure with large num­bers of people (voters or potential voters and candidates or potential candi­dates), usually across a substantial number of elections. 102 For example, the measurement of impact may factor elections over several cycles for every seat in a legislative chamber or congressional delegation. Cf. League of United Latin Am. Citizens v. Perry, 548 U.S. 399, 436–37 (2006) (holding “proportionality” in section 2 challenge to statewide redistricting is to be assessed on statewide basis). Under such condi­tions, large racial disparities are very unlikely to occur due to chance alone. 103 This follows from a statistical theorem called the “law of averages” or “law of large numbers.” For an introduction, see David Freedman et al., Statistics 273–84 (4th ed. 2007) (providing mathematical explanations and exercises to demonstrate law of averages). A showing of racially disparate impact will in most cases estab­lish by implication that there exists some set of race-correlated “so­cial con­ditions” (e.g., wealth, education, employment, political interest, church attend­ance, car ownership, newspaper readership, etc.) that ac­count for the disparity.

In point of fact, however, most courts addressing “social and histori­cal conditions” in a section 2 case focus not on socioeconomics generally, but rather on subjective race discrimination and its legacy—precisely what the courts ought to focus on if the purpose of the showing is to iden­tify risk factors for unconstitutional race discrimination in the elec­toral process. 104 Many such factors were laid out in an influential Senate Committee report (al­though not characterized by the Committee as “risk factors” for constitutional violations). See Katz et al., supra note 30, at 648–50 (“The Senate Report identified several factors, now known as ‘the Senate Factors,’ for courts to use when assessing whether a particular prac­tice or procedure results in prohibited discrimination in violation of Section 2.”). For a particularly clear judicial statement interpreting the social-and-historical-conditions re­quirement much as we suggest, see League of Women Voters of N.C. v. North Carolina, 769 F.3d 224, 240 (4th Cir. 2014) (“[The] burden must in part be caused by or linked to social and historical conditions that have or currently produce discrimination against members of the protected class.” (emphasis added) (citations omitted) (internal quotation marks omitted)). Typical considerations include Jim Crow history, housing segregation, racially polarized voting (often assumed to manifest discrim­ination against minority candidates), public and private discrimination in employment and education, racial campaign appeals, etc. 105 See generally Katz et al., supra note 30, at 661–730 (discussing judicial treatment of “Senate Report” factors). For recent exemplars, see, e.g., League of Women Voters, 769 F.3d at 238–47 (emphasizing history of discrimination); Ohio State Conference of NAACP v. Husted, 768 F.3d 524, 556–59 (6th Cir. 2014) (emphasizing discrimination in education and employment as well as in electoral process). These are risk factors for unconstitutional state action in the electoral process inso­far as they bespeak a world in which race discrimination is pervasive.

Some courts go further and expressly read section 2 as requiring proof of race-discriminatory “causation”—evidence that the injury for which plaintiffs seek redress resulted from subjective discrimination by conventional state actors or by voters. 106 See generally Hebert et al., Realist’s Guide, supra note 30, at 57–58 (“[A] num­ber of courts have held that a plaintiff cannot win a Section 2 case where racially polarized voting is caused by reasons other than racial animus.”); Greiner, Causal Inference, supra note 63, at 590–97 (stating “prevailing view” is that “section 2 does contemplate . . . causal inquiry”); Katz et al., supra note 30, at 670–73 (“Courts in nine judicial circuits now ex­pressly or implicitly incorporate causation.”). For a rare exception, see United States v. Blaine County, 363 F.3d 897, 912 (9th Cir. 2004) (“Requiring proof of discriminatory mo­tives among white voters in Blaine County would be divisive and would place an impossible burden on the plaintiffs.”). On the particulars of this causa­tion requirement, the courts are all over the map. Some courts say it is a necessary element of a section 2 case. 107 See, e.g., Smith v. Salt River Project Agric. Improvement & Power Dist., 109 F.3d 586, 595 (9th Cir. 1997) (requiring section 2 plaintiffs to demonstrate causal connection between challenged voting practice and prohibited discriminatory result); Uno v. City of Holyoke, 72 F.3d 973, 983 (1st Cir. 1995) (“[T]he ultimate burden of persuading the fact­finder that the voting patterns were engendered by race rests with the plaintiffs.”); S. Christian Leadership Conference of Ala. v. Sessions, 56 F.3d 1281, 1292 (11th Cir. 1995) (finding lack of “discriminatory purposes” in place requirements and lack of “discrimina­tory intent” in size of judicial election districts); Nipper v. Smith, 39 F.3d 1494, 1515 (11th Cir. 1994) (en banc) (plurality opinion) (Tjoflat, C.J.) (“[I]f the evidence shows . . . that the community is not motivated by racial bias in its voting patterns, then a case of vote dilution has not been made.”); League of United Latin Am. Citizens, Council No. 4434 v. Clements, 999 F.2d 831, 862 (5th Cir. 1993) (en banc) (“[A]n inquiry into the causes un­derlying polarized voting is appropriate . . . .”); Mallory v. Ohio, 38 F. Supp. 2d 525, 575–76 (S.D. Ohio 1997) (stating plaintiffs must prove “‘deprivation of the minority group’s right to equal participation must be on account of a classification, decision, or practice that depends on race or color’” (quoting Nipper, 39 F.3d at 1515)), aff’d, 173 F.3d 377 (6th Cir. 1999). Others say it is just one factor among many to be weighed. 108 See, e.g., Goosby v. Town of Hempstead, 180 F.3d 476, 493–503 (2d Cir. 1999) (holding causation is relevant factor to be considered in totality of the circumstances in­quiry); see also Bone Shirt v. Hazeltine, 336 F. Supp. 2d 976, 1008–38 (D.S.D. 2004) (rec­ognizing causation “may be relevant” at totality of circumstances stage of section 2 case). Some courts seem to infer causation from Jim Crow history. 109 See, e.g., Miss. State Chapter, Operation Push v. Mabus, 932 F.2d 400, 405 (5th Cir. 1991) (affirming district court finding of section 2 violation based on disparity in voter registration rates “coupled with a history of discriminatory voter registration procedures”). For additional discussion and examples, see Katz et al., supra note 30, at 675–77 (describ­ing courts’ findings about the “nature, frequency, and recentness” of discrimination). To be clear, the courts in such cases are not necessarily asserting that the history means that racially polarized voting (or a particular state action) has been “caused” by racial discrimi­nation. Rather, their position seems to be that the history establishes a “totality of circum­stances” connection between present-day disparate impacts and intentional discrimination by state actors. Others insist on evidence that current voting pat­terns or actions by government officials manifest subjective discrimina­tion against minority candidates. 110 See, e.g., Uno, 72 F.3d at 983 (“If proven, these preconditions give rise to an in­ference that racial bias is operating through the me­dium of the targeted electoral struc­ture to impair minority political opportunities.”). Still other courts rebuttably presume sub­jective race discrimination from racially polarized voting in biracial elec­tions. 111 See, e.g., Goosby, 180 F.3d at 502–03 (Leval, J., concurring) (“Proof of [Gingles] fac­tors sufficiently supports an inference that race may have been a motivating factor to justify imposing on defendants the burden to prove that the regular defeat of minority preferred candidates is not the result of race-based intent . . . .”); Milwaukee Branch of the NAACP v. Thompson, 116 F.3d 1194, 1199 (7th Cir. 1997) (“Proving discriminatory intent is not part of the plaintiffs’ case under § 2.”); Uno, 72 F.3d at 983 (“The resultant infer­ence is not immutable, but it is strong; it will endure unless and until the defendant ad­duces credible evidence tending to prove that detected voting patterns can most logically be explained by factors unconnected to the intersection of race with the electoral sys­tem.”). The rebuttable presumption is also present, though more implicit, in circuits where “cau­sation” is considered at the totality of circumstances stage (rather than as part of the Gingles inquiry into bloc voting) and the Gingles factors are collectively regarded as establishing a presumption of liability. See, e.g., Teague v. Attala County, 92 F.3d 283, 294 (5th Cir. 1996) (holding plaintiffs need not prove “causal connection” between past dis­crimination and depressed minority participation); Jenkins v. Red Clay Consol. Sch. Dist. Bd. of Educ., 4 F.3d 1103, 1135 (3d Cir. 1993) (holding district court must “explain with particularity” finding for defendant in section 2 case where “white voters voting as a bloc . . . defeat the candidate of choice of a politically cohesive minority group”); see also NAACP v. City of Niagara Falls, 65 F.3d 1002, 1019 n.21 (2d Cir. 1995) (citing Jenkins, 4 F.3d at 1135); Clark v. Calhoun County, 21 F.3d 92, 97 (5th Cir. 1994) (same). Judicial treatment of the section 2 causation “requirement” is so varied and inconsistent that leading scholars don’t even agree whether the requirement has practical bite. Ellen Katz characterizes it as a signifi­cant barrier to section 2 claims. 112 Katz et al., supra note 30, at 671–72 & n.145 (citing thirteen cases and noting “[p]roving the linkage is difficult . . . and numerous lawsuits have held that plaintiffs failed to meet their burden . . . on this point”). Jim Greiner thinks it is a nomi­nal re­quirement only, regularly ignored in practice. 113 Greiner, Causal Inference, supra note 63, at 591 (“[T]he causal inquiry . . . ap­pears to matter little in actual cases unless the factual record demonstrates that candidates of minority race have enjoyed some measure of electoral success.”); Greiner, Re-Solidifying Racial Bloc Voting, supra note 69, at 459–60 (characterizing current burden-shifting prac­tices as depriving causation requirement of practical bite).

In any event, our aim is not to extract from the case law some major­ity or plurality view of the causation or social-and-historical-conditions requirement. Rather, the intention is to gloss it in a manner that a broad range of judges could accept, including those—such as the median Justice on the current Supreme Court—who are leery of equalizing, race-conscious state action except as a remedy for subjective race discrimina­tion. As noted above 114 See supra notes 98–100 and accompanying text (noting “social and historical con­ditions” requirement may maintain nexus between liability under results test and ac­tual or threatened constitutional violations). and explained at length elsewhere, the “risk fac­tors” interpretation effectively reconciles the results test of section 2 with the jurisprudential premises of the conservative center. 115 See Elmendorf, Making Sense of Section 2, supra note 23, at 417–47 (developing interpretation of section 2 grounded in statute’s “legislative history and constitutional context”). The point of the showing is to establish that the relief the plaintiff seeks is reasonably well tailored to remedy or prevent constitutional violations. Plaintiffs need not prove that constitutional violations in fact occurred, or that they necessarily will occur in the future. But plaintiffs must establish a significant likelihood or risk of unconstitutional race discrimination in the electoral process. Accordingly, this Article will refer to this element of the plaintiff’s case not as the “social and historical conditions” require­ment or the “causation” requirement, but rather as the “constitutional risk” or “risk factors” requirement.

One might suppose that constitutional risk could be established, at least in the Deep South, merely by invoking Jim Crow history. Remedies under section 2 can certainly be characterized as remedying the linger­ing effects of generations of unchecked discriminatory state action in the South. 116 See Pamela S. Karlan, Two Section Twos and Two Section Fives: Voting Rights and Remedies After Flores, 39 Wm. & Mary L. Rev. 725, 738–40 (1998) (discussing reme­dies to racial bloc voting that “dampen the present effects of . . . past external discrimination”). Numerous lower courts have in fact emphasized Jim Crow his­tory as part of the “social and historical conditions” inquiry. 117 See Katz et al., supra note 30, at 676 (noting “[m]any courts” evaluating social and historical conditions discussed discrimination from nineteenth and early twentieth cen­turies, including “literacy tests, grandfather clauses, poll taxes, white primaries, . . . voter registration requirements, . . . [and] state laws mandating segregation”). But in Shelby County v. Holder, the Supreme Court denied the legitimacy of his­tory as a basis for subjecting state and local governments to preclearance under section 5. 118 See Joel Heller, Shelby County and the End of History, 44 U. Mem. L. Rev. 357, 359–60 (2013) (noting Court’s rejection of section 5 formula relying on historical data); see also Shelby County v. Holder, 133 S. Ct. 2612, 2629 (2013) (holding Congress “cannot rely simply on the past” and must use “current data” for coverage formula). Shelby County asks instead for evidence of current race discrimination, not historical discrimination; the Court was unwilling to assume that state and local governments that engaged in de jure discrim­ination half a century ago are today more likely than other subnational political units to discriminate unconstitutionally. 119 133 S. Ct. at 2625–31. Shelby County is of a piece with many other Supreme Court decisions setting time limits on remedies for past de jure discrimination. 120 See, e.g., Bd. of Educ. v. Dowell, 498 U.S. 237, 247 (1991) (“From the very first, federal supervision of local school systems was intended as a temporary measure to remedy past discrimination.”). As Gary Orfield and Chungmei Lee observe, “The basic position of the Court during the past 16 years has been that the constitutional violations arising from a history of segregation and inequality, when proved, justify race conscious remedies but only for a limited time.” Gary Orfield & Chungmei Lee, Historic Reversals, Accelerating Resegregation, and the Need for New Integration Strategies, Report of the Civil Rights Project/Proyecto Derechos Civiles, UCLA 7 (2007), http://civilrightsproject.ucla.edu/research/k-12-education/integration-and-diversity/historic-reversals-accelerating-resegregation-and-the-need-for-new-integration-strategies-1/orfield-historic-reversals-accelerating.pdf [http://perma.cc/6G5B-GG85]; cf. Grutter v. Bollinger, 539 U.S. 306, 343 (2003) (“We expect that 25 years from now, the use of racial preferences will no longer be necessary to further the interest approved today.”). It is therefore doubtful that the Court will treat purely historical evidence as sufficient to establish a current risk of constitutional violations in section 2 cases. And it is equally unlikely that the Court will view results-oriented remedies under section 2 as a “congruent” response to constitutional violations that took place long ago. 121 Cf. Shelby County, 133 S. Ct. at 2624–29 (recounting “dramatic[]” improvements in minority political participation in South over last fifty years); Nw. Austin Mun. Util. Dist. No. One v. Holder, 557 U.S. 193, 202 (2009) (“Things have changed in the South.”).

Plaintiffs will be on much stronger ground if they can make the con­sti­tutional risk showing using current data, focusing on the present-day risk of unconstitutional race discrimination in the electoral process. History must be downplayed. Working from these premises, the balance of this Part sketches a set of rebuttable presumptions that could be used to im­plement the constitutional-risk and disparate-impact prongs of section

B. Presumptions About “Risk Factors” for Constitutional Violations

The essential ingredients for a workable showing of “constitutional risk” under section 2 are a facially convincing theory about risk factors for unconstitutional race discrimination and an empirical method for ascertaining the relative severity of those risk factors across jurisdictions using current data. The theory has to be facially convincing because the relevant constitutional violations are difficult to observe. Social scientists can track outcomes—minority registration and turnout, the election of minority candidates, etc.—but as presently interpreted, the Fourteenth and Fifteenth Amendments are indifferent to outcomes as such. What makes a racially disparate outcome unconstitutional is not the extent of the disparity but whether it results from subjective discrimination, and such discrimination is hard to detect.

Accordingly, this section suggests a pair of constitutional risk pre­sump­tions that speak to motive and propensities for disparate treatment. One is grounded on current racial attitudes, the other on elected offi­cials’ incentive to use race as a screening device to effectuate political discrimination. 122 These are not the only plausible options. It may also be feasible to estimate geo­graphic variation in disparate treatment of minority-race candidates using survey experi­ments. See Marisa A. Abrajano, Christopher S. Elmendorf & Kevin M. Quinn, Using Experiments to Estimate Racially Polarized Voting 27–35 (Univ. Cal. Davis Legal Studies Research Paper, No. 419, 2015), http://papers.ssrn.com/sol3/papers.cfm?abstract_id=‌256
9982 (on file with the Columbia Law Review) [hereinafter Abrajano et al., Using Experiments] (demonstrating technical feasibility of experiment with non-representative sample survey).

1. Racial Attitudes and Beliefs. 123 Portions of this section previously appeared in Elmendorf & Spencer, Preclearance, supra note 19, at 1137–38. — There may be no surer proposi­tion in constitutional law than that state action motivated by racial stereo­types or racial animus offends the Equal Protection Clause. In jurisdic­tions where majority-group voters subscribe to exceptionally dim views of a minority group, it is reasonable to presume that the regular defeat of minority candidates is due at least in part to constitutionally prohibited motives. It is also plausible to suppose that in these communities, the adop­tion or maintenance of electoral arrangements that disadvantage minori­ties tends to occur in part because of motives or beliefs that the Constitution disallows as the basis for state action. Locally elected offi­cials are selected from and by the residents, and if majority-race residents denigrate the mi­nority, these officials will probably be rewarded (or at least not pun­ished) at the ballot box if they suppress minority political participation or representation. 124 Note also that vote dilution remedies can be conceptualized as a response to un­constitutional state action by the electorate, see Elmendorf, Making Sense of Section 2, supra note 23, at 430–47.

To be sure, it doesn’t follow from the existence of negative racial at­ti­tudes on the part of the white majority that minority candidates or mi­nority voters will in fact suffer disparate-treatment discrimination. In­deed, the question of whether racial attitudes “cause” disparate treat­ment is, for methodological purists, unanswerable. Like her race itself, 125 See D. James Greiner & Donald B. Rubin, Causal Effects of Perceived Immutable Characteristics, 93 Rev. of Econ. & Stat. 775, 776 (2011) (“[A]ttributes are not subject to change by intervention.”). a person’s racial attitudes cannot be manipulated by researchers, and with­out manipulation of the supposed “cause” there is no way to deter­mine with certainty the effects of that cause. 126 See id. at 775 (noting impossibility of manipulating immutable traits “in a way analogous to administering a treatment in a randomized experiment”). On the centrality of manipulation/randomization to causal inference, see Joshua D. Angrist & Jörn-Steffen Pischke, Mostly Harmless Econometrics: An Empiricist’s Companion 5 (2008). Moreover, whatever “treat­ments” (life experiences) may cause the development of racial stereo­types probably cause many other things as well. So even if a treatment were shown to cause both the development of negative stereotypes of minorities and a reluctance to vote for minority candidates, it would not be clear that the negative stereotypes were responsible for the subject’s lack of support for minority candidates. 127 In all social science applications, the barriers to inference about causal pathways tend to be formidable. See Donald P. Green, Shang E. Ha & John G. Bullock, Enough Already About Black Box Experiments: Studying Mediation Is More Difficult than Most Scholars Suppose, 628 Annals Am. Acad. Pol. & Soc. Sci. 200, 200 (2010) (arguing statisti­cal methods used to study mediation are flawed and require strong assumptions to answer questions).

But these fine points about causal inference miss a more basic social reality: Racial attitudes are conventionally understood to motivate behav­ior. Bigots would not be castigated if bigotry were believed to be merely a set of attitudes unconnected to behavior. Law is a practical endeavor. Sometimes a social convention about causation is enough. 128 As Justice Souter once observed about constitutional judicial review, “The quan­tum of empirical evidence needed to satisfy heightened judicial scrutiny of legislative judg­ments will vary up or down with the novelty and plausibility of the justification raised.” Nixon v. Shrink Mo. Gov’t PAC, 528 U.S. 377, 391 (2000). Consider also the torts doctrine of res ipsa loquitur, which clearly rests on generally shared beliefs (rather than data) about how things work in the world. See Jane Stapleton, Law, Causation, and Common Sense, 8 Oxford J. Legal Stud. 111, 112 (1988) (reviewing H.L.A. Hart’s work on role of “common sense” in legal treatments of causation). Subject to two provisos, the “constitutional risk” element of a section 2 claim can be deemed (presumptively) satisfied if plaintiffs show that majority-race citi­zens harbor negative attitudes about the minority. 129 Note that although conservative judges have generally rejected the idea that there is a compelling interest in remedying societal discrimination, Congress’s power to remedy societal discrimination may be considerably greater when such discrimination proximately affects voting and election outcomes. See Elmendorf, Making Sense of Section 2, supra note 23, at 430–36 (arguing electorate as a whole is state actor when it performs “public function” of putting in office officials who exercise coercive power of state); Ellen D. Katz, Reinforcing Representation: Congressional Power to Enforce the Fourteenth and Fifteenth Amendments in the Rehnquist and Waite Courts, 101 Mich. L. Rev. 2341, 2355–60 (2003) (not­ing Waite Court left room for Congress to attack race dis­crimination insofar as it pre­vented blacks from voting). Notably, the conservative judges who have read a race-discriminatory “causation” requirement into section 2 have accepted that race discrimination by the elec­torate furnishes the necessary causal link. See, e.g., Nipper v. Smith, 39 F.3d 1494, 1515–24 (11th Cir. 1994) (en banc) (plurality opinion) (Tjoflat, C.J.) (concluding plaintiff may prove discrimination by showing “objective factors that, under the totality of the circumstances, show the exclusion of the minority group from meaningful access to the political process due to the interaction of racial bias in the community with the challenged voting scheme”); League of United Latin Am. Citizens, Council No. 4434 v. Clements, 999 F.2d 831, 862 (5th Cir. 1994) (en banc) (stating plain­tiffs’ causation requirement satisfied by showing “fail­ure to elect representatives of their choice is attributable to white bloc voting rooted in racial considerations”).

The first proviso is that the measure of racial attitudes must correlate with political behavior or preferences among white (majority-race) voters. 130 Plaintiffs might satisfy this proviso by showing that more prejudiced whites (per the plaintiffs’ measure of prejudice) are less supportive than other whites of minority candi­dates compared to similar white candidates, or less supportive of certain policies when the policy beneficiaries are portrayed as plaintiff-race rather than white. The necessary show­ing could be made with observational or experimental data. Of course, the showing will not be causal in that racial attitudes themselves cannot be randomized. The showing would be cor­relational and therefore suggestive only. In section IV.B, infra, we satisfy the proviso by showing that our measure of prejudice predicts vote choice in elections con­tested by Barack Obama. See infra text accompanying notes 298–300. Section 2 is ultimately concerned with equal political opportunity. 131 See 52 U.S.C. 10301(b) (Supp. II 2015) (“A violation . . . is established if, based on the totality of the circumstances, it is shown that . . . members of a [protected] class of citizens . . . have less opportunity than other members of the electorate to participate in the political process and to elect representatives of their choice.” (emphasis added)). If whites’ racial attitudes do not correlate with political behavior, there is little ground for presuming that minority-preferred candidates or policies would have fared better but for white prejudice. By contrast, if whites’ racial attitudes are strongly associated with, for example, white support for minority-race can­di­dates, it makes sense to guard against the risk of discrimination even if the causal effect of racial attitudes on vote choice cannot be established. Just as a strong correlation between cholesterol levels or obesity, on the one hand, and heart disease, on the other, would justify some precaution­ary medical or dietary interventions, so too may correlational evidence justify legal interventions. 132 Thanks to Kevin Quinn for suggesting this analogy. The case for relying on cor­rela­tional evidence in the legal setting considered here is, on its face, stronger than the case for relying on correlational evidence in the heart-disease example. In the legal set­ting, the causal mechanism (linking racial attitudes to behavior) is knowable to some ex­tent through introspection or everyday social interaction, whereas in the medical setting intuition is probably not a good guide for laypersons.

The second proviso is that the racial-attitude measure must capture an at­titude or belief that the Constitution disallows as the basis for state action. Many political scientists have sought to quantify what they call “racial re­sent­ment” or “modern” racism, using survey questions that ask about support for federal welfare programs, affirmative action, and interven­tions by the “government in Washington” to improve the social and eco­nomic welfare of blacks. 133 See, e.g., Donald R. Kinder & Lynn M. Sanders, Divided by Color: Racial Politics and Democratic Ideals 15–18 (1996) (analyzing national surveys on equal opportunity, racially based federal assistance, and affirmative action to assess public views of racial mat­ters); Christopher Tarman & David O. Sears, The Conceptualization and Measurement of Symbolic Racism, 67 J. Pol. 731, 736–37 (2005) (describing use of surveys to test theory of symbolic racism). Racists no doubt give predictable answers to these questions, but there is nothing unconstitutional about predicating state action on the belief that federal welfare programs are a waste of money, or the view that affirmative action and efforts by the “government in Washington” to improve the welfare of blacks have been counterpro­ductive. Because of this, it is tenuous to infer a likelihood of unconstitu­tional race discrim­ination from the prevalence of “racial resentment.”

However, as we have shown elsewhere, conventional survey-based measures of racial stereotyping, which tap perceptions of racial differences in work ethic, intelligence, and trustworthiness, easily satisfy both of the provisos. 134 See Elmendorf & Spencer, Preclearance, supra note 19, at 1142–55 (explaining why survey data can capture racial attitudes correlated with political attitudes Constitution disallows as basis for state action). The conventional measures tap perceptions of racial dif­ferences in work ethic, intelligence, trustworthiness, and the like. These measures aren’t perfect—some respondents may not understand their own biases and others may not re­port their biases truthfully. The conventional measures may also be weaker for non-black minorities. But until better measures are produced, the conventional measures should suffice. Accordingly, the “constitutional risk” question may be re­solved (presumptively) by examining whether white voting-age citizens in the defendant jurisdiction subscribe to substantially negative stereotypes of the minority group. The courts, perhaps aided by DOJ, will need to define a quantitative benchmark for what constitutes legally significant racial stereotyping. 135 See infra section III.B (discussing need for courts to coordinate common presump­tions and develop “legally significant” racial stereotyping standard for section 2 litigation). Once this benchmark has been set, the question of whether a particular jurisdiction falls above or below it can be answered using national survey data and multilevel statistical modeling.

2. Racial Polarization in Partisanship. — Instead of (or in addition to) focusing on discrimination in the electorate, courts might ground “con­stitutional risk” presumptions on whether public officials have cause to discriminate against the racial minority so as to perpetuate their hold on power. Under certain conditions, those in power may benefit from sup­pressing minority participation irrespective of whether the officials (or majority-race voters) negatively stereotype the character or abilities of the minority.

This political incentive to discriminate most clearly arises when there is a strong correlation between voters’ race and their reliability as parti­san voters (or as consistent voters for any other established political fac­tion). Blacks, for example, are reliable Democratic voters. 136 For some evidence of this and a model of associated political incentives, see Cox & Holden, supra note 28, at 564–79. So when Republicans hold the reins of power, they have political incentives to di­minish black turnout. In recognition of this incentive, courts might deem the constitutional-risk requirement presumptively satisfied in cases brought by black voters against Republican-enacted voting requirements or redis­tricting plans so long as the plaintiffs show disparate impact and establish that the political incentive to discriminate holds in the defendant juris­diction, not just in the nation generally. 137 Cf. Growe v. Emison, 507 U.S. 25, 41–42 (1993) (rejecting that section 2 cases may be resolved on basis of what is typical nationally, rather than what is true in defendant jurisdiction).

To be sure, reasonable people may disagree about the propriety of inferring race discrimination, even presumptively, from “political incen­tives plus disparate impact.” Given present political alignments, the political-incentives presumption would tend to hobble Republican but not Democratic efforts to adjust electoral ground rules for partisan advantage. 138 In recent years, Republicans have promoted reforms that would make it harder for low-income and young people (groups disproportionately comprised of racial minori­ties) to vote, whereas Democrats have sought to expand turnout among the same de­mographics. See Richard L. Hasen, The Voting Wars: From Florida 2000 to the Next Election Meltdown 163–67 (2012) (describing recent debates between Republicans and Democrats over voter fraud and voter-ID election laws as battles of “access versus integrity”). This may make the presumption too politically fraught for the courts to adopt. 139 An intentionalist judge might also speculate that the median member of the coali­tion that enacted section 2’s results test would not have supported the political-incentives presumption. The results test emerged from a bipartisan compromise. See Boyd & Markman, supra note 67, at 1414–20 (detailing role of Senator Bob Dole in facilitating compromise on 1982 Amendments to VRA).

Another objection is that the political incentives presumption would really capture incentives to discriminate on the basis of partisanship, not race. Courts have long struggled to distinguish racial from political dis­crimination in section 2 and equal protection cases. 140 See Greiner, Causal Inference, supra note 63, at 593–95 (arguing courts make in­correct assumptions about impact of race on political processes); Richard L. Hasen, Race or Party? How Courts Should Think About Republican Efforts to Make It Harder to Vote in North Carolina and Elsewhere, 127 Harv. L. Rev. Forum 58, 67–70 (2014), http://cdn.harvardlawreview.org/wp-content/uploads/pdfs/forvol127_hasen.pdf [http://perma.cc/K2T5-GX4B] (discussing cases in which Court could not separate unconstitu­tional gerrymandering based on race from permissible redistricting based on party con­siderations); Pamela S. Karlan & Daryl J. Levinson, Why Voting Is Different, 84 Calif. L. Rev. 1201, 1220–27 (1996) (arguing courts’ difficulty disaggregating race and politics in section 2 cases is caused by race and political affiliation being “substantially correlated”); Richard H. Pildes, Is Voting-Rights Law Now at War with Itself? Social Science and Voting Rights in the 1990s, 80 N.C. L. Rev. 1517, 1565–67 (2002) [hereinafter Pildes, Voting-Rights Law] (“The Supreme Court’s key engagement with these amendments reflects deep divisions on whether separating racial from partisan reasons for differences in pref­erences between black and white voters is appropriate.”). Political discrim­ination is generally regarded as constitutionally innocuous, whereas race discrimination is deemed invidious. 141 See, e.g., Easley v. Cromartie, 532 U.S. 234, 241 (2001) (“We must determine whether there is adequate support for the District Court’s key findings, particularly the ul­timate finding that the legislature’s motive was predomi­nantly racial, not political.”).

Despite its roots in some important election law cases, 142 Id. the partisanship-not-race objection is hard to square with broader constitutional princi­ples. The Equal Protection Clause prohibits state actors from classifying persons by race and subjecting them to disparate treatment, unless doing so advances a compelling state interest that cannot be protected using race-neutral means. 143 See, e.g., Johnson v. California, 543 U.S. 499, 507 (2005) (applying strict scrutiny to California’s practice of segregating inmates by race during sixty-day evaluation period, notwithstanding undisputed evidence concerning violent prison gangs organized along racial lines); cf. J.E.B. v. Alabama ex rel. T.B., 511 U.S. 127, 139 n.11 (1994) (stating gender-based classifications utilizing stereotypes violate Equal Protection Clause “even when some statistical support can be conjured up for the generalization”). Racial animus and ugly stereotypes are not pre­requisites for an equal protection violation. 144 Judge Kozinksi explains the point nicely:
The lay reader might wonder if there can be intentional discrimina­tion with­out an in­vidious motive. Indeed there can. A simple example may help illus­trate the point. Assume you are an anglo homeowner who lives in an all-white neighborhood. Suppose, also, that you harbor no ill feelings toward minorities. Suppose further, however, that some of your neighbors persuade you that having an integrated neighborhood would lower property values and that you stand to lose a lot of money on your home. On the basis of that belief, you join a pact not to sell your house to minorities. Have you engaged in intentional racial and eth­nic discrim­ination? Of course you have. Your personal feelings toward minorities don’t matter; what matters is that you intentionally took actions calcu­lated to keep them out of your neighborhood.
Garza v. County of Los Angeles, 918 F.2d 763, 778 n.1 (1990) (Kozinski, J., concurring and dissenting in part).

If the correlation between race and partisan voting behavior is ex­tremely high, politically motivated state actors will have strong incentives to classify and target voters on the basis of their race. Race is generally easy to observe. Consistent partisan voting behavior is much harder to observe, for the ballot is secret and citizens do not wear their voting his­tory on their sleeves. Because race is more readily observed than reliable partisanship, elites seeking partisan political advantage have incentives to target voters on the basis of their race.

That said, the Supreme Court has been wary about applying the anti­stereotyping logic of equal protection doctrine in cases about political discrimination. In racial gerrymandering cases, for example, the Court has crafted decision rules that make it very difficult to challenge state ac­tions that classify voters by race if the action can be explained as a par­ti­san maneuver and the racial classification is not facially evident. 145 See Easley, 532 U.S. at 258 (holding plaintiffs challenging electoral district on ground race “predominated” over other considerations in its design must show defendant could have achieved its political objectives equally well using more racially heterogeneous districts); see also Elmendorf & Spencer, Preclearance, supra note 19, at 1134–35 (discuss­ing Easley). It is perhaps a sign of changes to come, however, that the Court’s most recent decision concerning a “racial predominance” claim, Alabama Legislative Black Caucus v. Alabama, 135 S. Ct. 1257 (2015), does not mention Easley’s requirement. But the Court has never denied the proposition that disparate treatment on the basis of race in the political sphere offends the Constitution’s equal pro­tection norm. 146 Indeed, in Voinovich v. Quilter, 507 U.S. 146 (1993), the Court seemed to ac­cept that racial vote dilution undertaken for partisan political reasons would be unconsti­tu­tional. Id. at 159–60 (holding evidence did not support district court’s finding that re­dis­tricting authority “sought to minimize the Democratic Party’s power by diluting minor­ity voting strength,” but not questioning district court’s premise that intentional race dis­crim­ination undertaken for partisan purposes is constitutionally proscribed).

It bears emphasis, finally, that the political incentives approach would not necessarily result in commonplace, Republican-preferred vot­ing requirements with a racially disparate impact being invalidated in jurisdictions with large minority populations and allowed to stand else­where. A defendant might rebut the inference of racial targeting by show­ing that voting restrictions similar to the one at issue are strongly backed by Republicans in states without a sizeable, heavily Democratic minority population. (Ordinary voter-ID requirements might survive; rollbacks of Sunday early voting in communities with politically mobilized black chur­ches probably would not.) Or defendants might show that the voting re­striction is well designed to advance important state interests, 147 One of the “totality of circumstances” factors that courts regularly consider in section 2 cases is the degree to which the challenged law is tenuous or advances important state interests. See, e.g., Hous. Lawyers’ Ass’n v. Texas, 501 U.S. 419, 426–27 (1991) (not­ing state interest in electing trial judges from districts coextensive with trial court’s jurisdic­tion “is a factor to be considered by the court in evaluating whether the evidence in a par­ticular case supports a finding [that this practice is] a vote dilution violation”); Katz et al., supra note 30, at 727–30 (reviewing case law in lower courts). or that the state made a good faith effort to monitor and curtail race discrimina­tion by administrators who implement the law.

By shifting the burden of persuasion to defendants, the courts sim­ply acknowledge that partisan motives do not merit the same presump­tion of legitimacy in jurisdictions where the partisan payoff to racial dis­crimination is exceptional.

To summarize, “constitutional risk” may be inferred, presumptively, from evidence of widespread negative stereotyping of the minority or (at least arguably) from evidence that the state actors responsible for the chal­lenged electoral arrangements had strong partisan incentives to discrimi­nate against the racial minority. The next question—and the subject of the next section—is whether analogous presumptions can be crafted for the disparate impact prong of section 2 cases.

C. Presumptions About Disparate Impact

As noted above, section 2 is concerned with two quite different sorts of impacts: racial disparities in opportunities to cast a valid, duly counted ballot (the issue in vote denial cases) and racial disparities in opportuni­ties to secure representation (the issue in vote dilution cases). Because dilution and denial can occur independently of one another, 148 The issue in typical dilution cases is whether the rules for translating votes into rep­resentation make it difficult for the minority community to secure representation—giv­en the distribution of political preferences in the electorate—even if minorities can vote with­out hindrance. See generally Heather K. Gerken, Understanding the Right to an Undiluted Vote, 114 Harv. L. Rev. 1663, 1671–81 (2001) [hereinafter Gerken, Understanding] (explain­ing concept of dilution as it developed in courts). different presumptions are needed for each class of cases.

1. Vote Dilution. — Though there is still ample room for disagree­ment about the meaning of disparate impact in dilution cases, 149 See generally Elmendorf et al., Racially Polarized Voting, supra note 60 (manu­script at 37–42) (contrasting four theories of racial vote dilution). most courts assess disparate impact by comparing the number of MODs—dis­tricts in which the minority community can elect its “candidates of choice”—with some benchmark conception of the appropriate or fair number of MODs. 150 Some courts also weigh opportunities for minority influence through channels other than the election of “candidates of choice.” See id. (manuscript at 38–40) (discuss­ing “coalitional breakdown” theory of vote dilution, and associated cases). Arguably the dominant approach is to treat “rough proportionality” as the fairness benchmark (equivalence between the minority’s population share and its share of electoral districts), but only insofar as proportionality can be achieved within a system of compact single-member districts drawn in accordance with traditional criteria. 151 See Hebert et al., Realist’s Guide, supra note 30, at 59 (“[P]roportionality has be­come an increasingly crucial issue in Section 2 cases.”); see also Katz et al., supra note 30, at 730–32 (noting in nearly all cases in which lower courts made finding on propor­tional­ity, liability question was resolved accordingly); cf. League of United Latin Am. Citizens v. Perry, 548 U.S. 399, 436 (2006) (beginning “totality of circumstances” inquiry by assessing proportionality, implying it is especially important factor to consider). Note, however, that LULAC’s ultimate finding of liability was not grounded on the lack of pro­portionality, and in a subsequent opinion Judge Easterbrook expressly rejected the pro­portionality bench­mark in favor of something like “the number of majority minority dis­tricts that probably would have been drawn by an automated redistricting algorithm fol­lowing traditional cri­teria.” Gonzalez v. City of Aurora, 535 F.3d 594, 599–600 (7th Cir. 2008). And it is certainly clear that the state has no obligation to achieve “proportionality” except insofar as it can be done by drawing reasonably compact, majority-minority districts. This follows from LULAC, which counts only “reasonably compact” majority-minority districts in the proportionality analysis, see 548 U.S. at 437–38, and which declares, “there is no § 2 right to a district that is not reasonably compact,” id. at 430 (citing Abrams v. Johnson, 521 U.S. 74, 111–12 (1997)).

To evaluate disparate impact, then, a court must establish whether the racial minority has distinct political preferences that set it apart from the majority group 152 See Thornburg v. Gingles, 478 U.S. 30, 47–51 (1986) (plurality opinion) (Brennan, J.) (“If the minority group is not politically cohesive, it cannot be said that the selection of a multimember electoral structure thwarts distinctive minority group inter­ests.”); Hebert et al., Realist’s Guide, supra note 30, at 48–50 (discussing meaning of mi­nority political cohesiveness). and if so, whether or to what extent the minority can elect its candidates of choice. (Absent some racial divergence in polit­ical preferences or interests, it does not make sense to speak of minority-race voters as a group having “candidates of choice.” 153 See Gerken, Understanding, supra note 148, at 1677–79 (considering impor­tance of group vote aggregation in influencing political process). )

It follows that evidence of racial polarization may come into play un­der the “disparate impact” as well as the “constitutional risk” prong of a dilution case—although evidence that a court deems sufficient under one prong may not satisfy the other. For example, a court might reasonably require extreme polarization in partisanship under the constitutional risk prong while holding that evidence of a statistically significant difference in policy preferences establishes preference polarization under the im­pact prong. 154 Subtle differences in different courts’ normative theories of vote dilution may also lead the courts to expect different kinds of showings of preference polarization. See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 42–49) (explain­ing conflicting implications of several normative theories of dilution).

Notice too that under the impact prong, evidence of racial diver­gence in political preferences or interests is clearly necessary but just as clearly not sufficient. Whether the minority community has the oppor­tunity to elect a roughly proportional number of its candidates of choice depends not only on the extent of preference polarization but also the geographic distribution of voters of each racial group and the rules for aggregating votes into outcomes. Courts need tools for determining whe­ther any particular legislative district is an MOD.

Presently, judges resolve the group-cohesion/polarization and opportunity-district questions using estimates of white and minority vot­ing patterns in recent elections in the defendant jurisdiction. 155 See generally id. (manuscript at 15–35) (reviewing and explaining judicial practice). These estimates are created by ecological inference from aggregate data—vote tallies and demographics at the precinct level. 156 See generally Greiner, Ecological Inference, supra note 78 (evaluating ecological inference methods used in assessing racial voting patterns). Generating the esti­mates is costly. Expert witnesses must retrieve several years’ worth of pre­cinct-level election results from county courthouses, digitize the data, merge it with demographic data from the Census, and then apply several differ­ent statistical tools for estimating the correlation between race and vote choice. 157 Interview with VRA Litigator, in Chi., Ill. (Apr. 6, 2014); Telephone Interview with Voting Rights Expert Witness (Jan. 3, 2014). Litigants then do battle over which elections are most “proba­tive” of group political cohesion and minority opportunity. 158 See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 24–27) (detailing selection and weighting of elections). Courts mud­dle through; the relevant legal doctrines give trial judges enormous dis­cretion but not much guidance about how to exercise it. 159 See id. (manuscript at 33) (“[T]here are no established quantitative cutoffs to dis­tinguish polarized from non-polarized communities, no clear-edge rules about which elec­tions to include in the polarization analysis and how to weight them, and . . . racial assump­tions are . . . baked into the statistical tools . . . .”).

This process could be greatly streamlined if the courts recognized re­buttable presumptions about racial-group cohesion and MODs, presump­tions whose application would depend on national survey data rather than local election results. The next two sections survey the possibilities, beginning with the question of preference/interest divergence and then moving on to the question of whether a district should be counted as an MOD.

a. Preference/Interest Divergence (Polarization). — Using national sur­vey data, there are several ways to answer the question of whether plain­tiffs belong to a racial community with distinct political interests or pref­erences, not shared by the racial majority: (1) base polarization deter­minations on voting-age citizens’ stated political preferences (preference polarization); 160 Whether the presumptions should reflect the political preferences of all voting-eli­gible citizens, or only registered voters or likely voters, is a question this Article does not resolve. (2) base polarization determinations on citizens’ inter­ests (interest polarization); or (3) base polarization determinations on the results of survey experiments (a variant on preference polarization). This section briefly describes the three approaches; Part IV reports origi­nal empirical results on preference polarization.

Existing national surveys contain a wealth of individual-level data about respondents’ policy positions, party identification, demographics, etc. With the aid of recently popularized statistical techniques, these data can be used to generate estimates of racially polarized preferences within small geographic units, such as congressional districts, state legislative districts, or counties.

Alternatively, census data can be used to establish differences be­tween racial groups in terms of economic position, health status, incar­ceration rates, and the like. This approach to the polarization inquiry presumes that people vote their interests rather than their principles, which is not always true. 161 Cf. Andrew Gelman et al., Red State, Blue State, Rich State, Poor State: Why Americans Vote the Way They Do 48–54 (expanded paperback ed. 2010) (showing system­atic regional differences in degree to which affluent people vote their economic interests); Eitan Hersh & Clayton Nall, The Primacy of Race in the Geography of Income-Based Voting: New Evidence from Public Voting Records, Am. J. Pol. Sci. (forthcoming 2015), http://www.stanford.edu/~nall/docs/cata9.6.pdf [http://perma.cc/P9RB-L6SX] (explor­ing geographic variation in correlation between income and partisan voting). But the interest-based approach has the ad­vantage of not relying on litigant-generated models to produce estimates of local public opinion, as the relevant data are available from the Census Bureau at the geographic scales needed for section 2 litigation. 162 The objective approach might be implemented with the types of factor analysis that Nick Stephanopoulos has used to measure the spatial heterogeneity of legislative dis­tricts. See Stephanopoulos, The South After Shelby County, supra note 40, at 94–101 (de­scribing refinements to factor analysis method of evaluating districts’ spatial diversity); Nicholas O. Stephanopoulos, Spatial Diversity, 125 Harv. L. Rev. 1903, 1907 (2012) (syn­thesizing socioeconomic data from Census to produce “single figure for each [congres­sional] district that shows . . . how spatially homogeneous or heterogeneous the district is”). The preference-based approach is, however, more in keeping with the exist­ing judicial focus on voter preferences, 163 See, e.g., United States v. Blaine County, 363 F.3d 897, 910 (9th Cir. 2004) (re­jecting any approach to polarization inquiry that would require judges to “second guess voters’ understanding of their own best interests”). But see League of United Latin Am. Citizens v. Perry, 548 U.S. 399, 430–35 (2006) (incorporating evidence of socioeconomic conditions into evaluation of district’s “compactness”). as well as recent empirical evi­dence about geographic variation in income-based voting and ontologi­cal commitments to free will. 164 Supra note 163.

Both preference- and interest-based approaches present challenges when racial groups polarize on some but not all issues, ideological di­men­sions, or interests. The responsible decisionmaker must decide how to weight the various indicators of cohesion or polarization. But this prob­lem—for purposes of a rebuttable presumption of cohesion or po­larization—is less vexing and less of a barrier to preliminary relief than the analogous problem, in a conventional racial polarization analysis, of deciding which elections to include in the analysis and how to weight them. 165 Courts struggle all the time with whether to include or how to weight voting data from white versus white elections, primary elections, and elections to governmental bodies other than the one at issue. See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 24–27, 29–31) (observing courts have “opted for loose guidelines” on types of elections to weigh in analysis, causing disagreement over probative value of “mono-racial,” primary, and “exogenous” elections).

One reason it is less vexing is that the rebuttable presumptions would be implemented using national survey data. This means that the same universe of issues and summary measures of preferences (or inter­ests) will be available in all section 2 cases. Once a circuit court decides that a particular measure suffices, either in general or for a particular type of governmental body, 166 One example would be measures of educational attainment used in school board election cases. subsequent section 2 cases can be brought in other states and localities using the very same measures. By contrast, courts answering the polarization question with data on vote shares usu­ally give the most weight to recent elections for the governmental body at issue in the case. 167 See Hebert et al., Realist’s Guide, supra note 30, at 54–55 (noting many courts have discounted, and some even refuse to consider, evidence of racial polarization in “ex­ogenous” elections, i.e., elections for governmental body other than one at issue in case). Each case therefore depends on sets of election re­sults unique to the case. The bottom line is that an evolving “common law” of racial polarization with respect to preferences or interests should provide more guidance regarding the likely outcome of the next case than has the common law of racial polarization with respect to vote shares in candidate elections. 168 This common law still might not provide enough guidance without a strong as­sist from DOJ. See infra notes 222–231 (discussing DOJ’s ability to induce judicial coor­dination by issuing section 2 interpretive rules). This has obvious implications for the availability of preliminary relief.

Second, because the presumption of racial polarization would be re­buttable, courts need not be perfectionist about the measure. A generic measure of ideology scaled from issue preferences (i.e., first dimension ideal points) arguably should suffice for most elections, 169 On scaling ideology from issue positions, see generally Joshua D. Clinton, Using Roll Call Estimates to Test Models of Politics, 15 Ann. Rev. Pol. Sci. 79 (2012) (reviewing methods of analyzing and interpreting roll call voting to determine ideology). For a treat­ment of some issues that arise when the same methods are used to scale ordinary citi­zens’ ideology, see generally Jeffrey B. Lewis & Chris Tausanovitch, Has Joint Scaling Solved the Achen Objection to Miller and Stokes? (Feb. 6, 2013) (unpublished manu­script), http://www.vanderbilt.edu/csdi/miller-stokes/05_MillerStokes_LewisTausanovitch.pdf [http://perma.cc/62Z6-W37K] (arguing methodologies for estimating legislator ideology from roll call voting are incompatible with public opinion polls). even though citizens with the same ideal points may have important disagreements on certain issues. 170 See David E. Broockman, An Artificial “Disconnect” 16–19 (Oct. 24, 2014) (un­published manuscript) (on file with the Columbia Law Review) (illustrating voters with aver­age responses across several issues appear moderate but provide extreme responses to sin­gle issues). Alternatively, judges could ask litigants to show the rela­tive importance that minority and white voters attach to different is­sues. 171 To fully implement this approach, the organizations that conduct large-scale na­tional surveys would have to be convinced to ask priorities questions alongside the issue-position questions. If DOJ asked for this information and provided some funding, we think the survey organizations would be more than happy to obtain it. Cohesion and polarization determinations could then be based on issue preferences weighted by their importance to minority voters. Of course, any court that continued to regard polarized voting in local elec­tions as particularly informative about group political cohesion could invite litigants to rebut the presumption with local voting data.

Perhaps the cleanest solution to the “which issues” problem is to base polarization determinations on preferences revealed through survey experiments. One of us shows in a working paper that this can be done by asking voters to choose between pairs of hypothetical candidates whose race and endorsements have been randomized. 172 See Abrajano et al., Using Experiments, supra note 122, at 13–18 (detailing study methodology). For a related idea, see generally Will Bullock, Kosuke Imai & Jacob N. Shapiro, Statistical Analysis of Endorsement Experiments: Measuring Support for Militant Groups in Pakistan, 19 Pol. Analysis 363 (2011) (randomizing group endorsement of poli­cy positions and using Bayesian hierarchical models to infer geographic variation in sup­port for endorsers).

Whichever measure one favors, the important point for present pur­poses is that it is feasible to create presumptions about within-group po­litical similarity and between-group political difference without reference to voting patterns in recent elections in the defendant jurisdiction. The viability of a section 2 claim need not depend on expensive expert witness anal­yses of local voting data, on statistically tenuous techniques of ecological in­ference, or on the happenstance of whether plaintiff-race candidates have recently run for office in the locale.

b. Minority Opportunity Districts. — The existence of significant po­larization in interests or preferences between white and minority com­munities does not necessarily mean that particular minority plaintiffs lack a realistic opportunity to elect their candidates of choice. If the minority community is large and if polarization is not too extreme, enough white voters may “cross over” and support minority-preferred candidates for the candidates to be electable. So polarization alone is not enough to es­tablish a presumption of dilution (disparate impact). The court needs some way of gauging whether particular districts are minority oppor­tunity districts and then comparing the share of MODs to the minority’s population share. 173 This assumes that what is being challenged is a system of single-member districts. If the plaintiffs were challenging at-large elections or multi-member districts, then what would be needed is a method for assessing minority opportunity under that system.

Historically the courts have assessed the likely “performance” of elec­toral districts with detailed inquiries into local political conditions. 174 See Hebert et al., Realist’s Guide, supra note 30, at 56–59 (explaining since Gingles courts have had to assess whether white bloc voting “usually [results in] defeat of the minority’s preferred candidate” and this inquiry requires courts to consider “a variety of factual circumstances” (internal quotation marks omitted)); Stephanopoulos, The South After Shelby County, supra note 40, at 80 (discussing predictive judgments about ability-to-elect under section 5). Into the mix go the results of past elections, the extent of racial polariza­tion, racial differences in voter eligibility and voter turnout rates, anecdo­tal testimony from local politicians, consultants, and interest groups, and more. 175 All of these factors figure into the “totality of circumstances” analysis of a section 2 case. See generally Katz et al., supra note 30, at 675–730 (discussing “Senate Factors”).

One straightforward way to simplify this inquiry is to presume that a district is an MOD if and only if the minority community composes at least 50% of the district’s citizen voting age population (CVAP). 176 CVAP estimates from the Census are less precise than estimates of the total vot­ing age population. See Nathaniel Persily, The Law of the Census: How to Count, What to Count, Whom to Count, and Where to Count Them, 32 Cardozo L. Rev. 755, 773–82 (2011) (arguing difficulties obtaining reliable citizenship data should lead Court to clarify total voting age population is appropriate metric for demonstrating size of minority com­munity). This imprecision does not concern us because the estimates would only be used to establish a rebuttable presumption and because estimation errors should to substantial degree wash out as CVAP estimates at the level of census blocks and tracts are aggregated to the level of legislative districts. Be­cause the Constitution prevents the government from erecting substan­tial barriers to registration and voting, 177 At the least, the Constitution prevents the government from erecting barriers that are substantial for some people but not for others. See Crawford v. Marion Cty. Election Bd., 553 U.S. 181, 198–200 (2008) (finding burdens imposed by Indiana voter-ID law not sufficiently substantial to invalidate statute partly because of mitigations to protect those burdened by law); Kramer v. Union Free Sch. Dist. No. 15, 395 U.S. 621, 627 (1969) (“[I]f a challenged state statute grants the right to vote to some bona fide residents of requisite age and citizenship and denies the franchise to others, the Court must determine whether the exclusions are necessary to promote a compelling state interest.”). it may be said that any district in which the minority community makes up at least half of the voting-eligible population is by definition an opportunity district. 178 See Jeffers v. Beebe, 895 F. Supp. 2d 920, 930–33 (E.D. Ark. 2012) (holding after Bartlett v. Strickland, 556 U.S. 1 (2009) (plurality opinion) (Kennedy, J.), any district which is majority-minority by voting-eligible citizens is minority opportunity district as mat­ter of law). Some of these dis­tricts may not “perform” for minority candidates owing to race-correlated differences in rates of voter registration, turnout, or information about the candidates, but since section 2 protects equality of opportunity rather than equality of results, 179 Johnson v. De Grandy, 512 U.S. 997, 1014 n.11 (1994). these differences arguably should be disregard­ed (unless they can be fairly attributed to race discrimination in violation of the Fourteenth and Fifteenth Amendments 180 See Salas v. Sw. Tex. Junior Coll. Dist., 964 F.2d 1542, 1556 (5th Cir. 1992) (not­ing protected class would be entitled to section 2 relief if turnout disparities were attribut­able to prior official discrimination); Theane Evangelis, Note, The Constitutionality of Compensating for Low Minority Voter Turnout in Districting, 77 N.Y.U. L. Rev. 796, 808 (2002) (“The Supreme Court’s current equal protection doctrine requires strict scrutiny review for race-conscious state policies, including excessively race-conscious districting.”). ).

In recent years, voting rights claimants have often argued that strict­ly majority-minority districts are not necessary to provide the “opportu­nity to elect” minority candidates of choice, and indeed that such dis­tricts may weaken the minority community’s overall political influence by wasting minority votes. 181 This issue is front and center in pending constitutional challenges to statewide re­districting maps in Alabama and North Carolina. See Ala. Legislative Black Caucus v. Alabama, 135 S. Ct. 1257, 1262, 1274 (2015) (vacating district court decision rejecting plaintiffs’ claim Alabama violated Fourteenth Amendment by drawing majority-minority districts not reasonably necessary to comply with VRA); see also Dickson v. Rucho, 135 S. Ct. 1843 (2015) (vacating similar decision of North Carolina Supreme Court for reconsid­eration in light of Alabama Legislative Black Caucus). Certainly there are more nuanced alternatives to the “50% CVAP” rule, such as presuming that a district is an MOD if the minority community composes a majority of the citizens in the dis­trict who prefer the major political party with the most support among the district’s voters; 182 Cf. League of United Latin Am. Citizens v. Perry, 548 U.S. 399, 485–86 (2006) (Souter, J., concurring in part and dissenting in part) (suggesting this standard as gloss on Gingles numerosity requirement). or classifying districts based on the joint distribu­tion of political and racial preferences in the district electorate. Space limitations preclude an adequate treatment of these alternatives here, but we plan to take them up in future work. 183 It might also be argued that the minority-opportunity-district presumption should account for race-correlated differences in voter turnout. Reasoning thus, some courts have applied a “65%” rather than a “50%” rule of thumb, see, e.g., Barnett v. City of Chicago, 141 F.3d 699, 702–03 (7th Cir. 1998) (“A black-majority ward, then, is one that is at least 65 percent black on a total-population basis . . . .”); Ketchum v. Byrne, 740 F.2d 1398, 1416 (7th Cir. 1984) (“Numerous courts have either specifically adopted or tacitly approved the use of this 65% figure.”), although this convention appears to have faded. See Pildes, Voting-Rights Law, supra note 140, at 1527–28 (“[B]y 1990, the 65% rule was considered exceptional.”). Evangelis, supra note 180, at 801, argues that adjusting for turn­out in this way is unconstitutional.

2. Vote Denial. — Crafting disparate-impact presumptions for vote de­nial cases is difficult. The range of election rules and practices that might be challenged on a denial theory is vast and so too the range of possible remedies.

The vote denial cases are quite new, and there is not yet much law on what constitutes a material, legally significant impact. 184 League of Women Voters of N.C. v. North Carolina, 769 F.3d 224, 239 (4th Cir. 2014) (“[T]here is a paucity of appellate case law evaluating the merits of Section 2 claims in the vote-denial context.”). The Seventh Circuit per Judge Easterbrook recently suggested that section 2 is vio­lated only if the challenged barrier to voting is pointless, 185 See Frank v. Walker, 768 F.3d 744, 753 (7th Cir. 2014) (“[U]nless Wisconsin makes it needlessly hard to get photo ID, it has not denied anything to any voter.”). facially dis­crimina­tory, 186 See id. (“[The Wisconsin voter-ID statute] extends to every citizen an equal op­por­tunity to get a photo ID.”); id. at 754 (“It is better to understand [the results test] as an equal-treatment requirement . . . than as an equal-outcome command . . . .”). or hard for diligent voters to comply with. 187 See id. at 752–53 (reviewing evidence of racial disparities in possession of qualify­ing photo ID and in documents required to obtain qualifying photo ID and concluding, “[a]lthough these findings document a disparate outcome, they do not show a ‘denial’ of anything by Wisconsin”). The district court in the same case characterized the burden very differently, focusing on ra­cial disparities in the incidence of compliance costs and the likeli­hood of a correlative disparity in rates of voter participation. 188 See Frank v. Walker, 17 F. Supp. 3d 837, 870–78 (E.D. Wis. 2014) (discussing ra­cial disparities in rates of possession of qualifying ID as well as documentation needed to obtain ID, linking these to socioeconomic disparities, and emphasizing “under the domi­nant framework used by scholars to study voter turnout, even small increases in the costs of voting can deter a person from voting”), rev’d, 768 F.3d 744. In the Sixth Circuit, a slight rollback in the days available for early voting was deemed sufficiently impactful to violate section 2. 189 See Ohio State Conference of NAACP v. Husted, 768 F.3d 524, 549–60 (6th Cir. 2014) (finding disparate impact and applying “Senate factors” to conclude early-voting re­strictions burdened black voters and are linked to “social and historical conditions”). Yet an earlier Sixth Circuit decision ruled out section 2 challenges to felon disenfranchise­ment laws on the theory that the felon had no one but himself to blame for getting disenfranchised. 190 Wesley v. Collins, 791 F.2d 1255, 1261–62 (6th Cir. 1986) (“Nor are felons disen­franchised because of an immutable characteristic, such as race, but rather because of their conscious decision to commit a criminal act for which they assume the risks of deten­tion and punishment.”); see also Ortiz v. City of Philadelphia, 28 F.3d 306, 315–17 (3d Cir. 1994) (rejecting challenge to law purging voters who have registered but failed to vote in past from voter rolls on ground that voter is at fault). An analogous voter-fault argument easily could have been used to dismiss the early-voting claim.

Because the law is inchoate and the cases diverse, rebuttable pre­sumptions for the disparate-impact prong of a vote denial case must be ventured tentatively. In that cautious spirit, we offer two ideas. First, ma­terial burdens on the franchise that are disproportionately borne by low-income citizens may be presumed to have a racially disparate impact in those jurisdictions where Census data show that minority-race citizens are substantially worse off economically than members of the majority group. Much remains to be worked out here. What measure of economic well-being should be used? What constitutes a “substantial” between-group disparity in economic well-being?

All that said, the idea that disparate impacts by class can serve as a proxy for perhaps harder-to-observe disparate racial impacts is already gaining traction in the lower courts. It is reflected in recent judicial opin­ions addressing photo-ID requirements for voting in Wisconsin and Texas, 191 See Veasey v. Perry, 71 F. Supp. 3d 627, 660–64 (S.D. Tex. 2014) (tying conclu­sions about racially disparate impact to finding that voter-ID law is more burdensome for poor and poor are disproportionately minority), aff’d in part, vacated in part, No. 14-41127, 2015 WL 4645642 (5th Cir. Aug. 5, 2015); Frank, 17 F. Supp. 3d at 870–79 (same); see also Texas v. Holder, 888 F. Supp. 2d 113, 140–43 (D.D.C. 2012) (denying preclearance to Texas voter-ID requirement, based on retrogressive effect inferred from correlation between race and poverty), vacated on other grounds, 133 S. Ct. 2886 (2013). and rollbacks in early voting and same-day voter registration in Ohio and North Carolina. 192 League of Women Voters of N.C. v. North Carolina, 769 F.3d 224, 246 (4th Cir. 2014) (using disparate impact viewed in context of social and historical conditions to find election reform law “textbook example of Section 2 vote denial”); Husted, 768 F.3d at 555–57 (linking disparate reductions in voting opportunities to social and historical conditions that “produce discrimination against African Americans”). The idea also draws support from the legis­lative history of section 2, which shows that Congress was quite con­cerned about socioeconomic disparities between racial groups manifest­ing as political inequalities. 193 S. Rep. No. 97-417, at 29 n.114 (1982) (“[D]isproportionate educational[,] em­ployment, income level and living conditions arising from past discrimination tend to depress minority political participation. Where these conditions are shown, and where the level of black participation . . . is depressed, plaintiffs need not prove any further causal nexus between their disparate socioeconomic status and the depressed . . . participation.” (citations omitted)).

The other presumption goes to the question of materiality: Any vot­ing requirement that has the demonstrable effect (compared to some feasible regulatory alternative) of skewing the racial/ethnic makeup of the population of actual voters, relative to the population of voting-eligible citizens, should be presumptively regarded as materially burdensome. Al­most certainly, this showing would have to be made using data from a range of jurisdictions, because a simple before-and-after comparison of voter participation in the defendant jurisdiction could be very misleading. Mi­nority turnout in a single jurisdiction may fluctuate for any number of reasons unrelated to the legal change. 194 Cf. Robert S. Erikson & Lorraine C. Minnite, Modeling Problems in the Voter Identification—Voter Turnout Debate, 8 Election L.J. 85, 96–98 (2009) (demonstrating difficulty of detecting small turnout effects). Courts should accept such evi­dence from other jurisdictions, so long as the demographics of the rele­vant minority populations are not grossly dissimilar.

Unlike the presumptions discussed above for constitutional risk, ra­cial polarization, and minority opportunity districts, the presumptions for vote-denial “impact” could not be implemented with national public opinion data. At best they are rough guidelines that may help courts as they struggle to produce a reasonably consistent body of law, notwith­stand­ing judges’ contrasting intuitions about the burdens that any “rea­sonable voter” ought to bear without complaint and the difficulty of de­tecting racially disparate impacts amidst all the fluctuations in voter regis­tration and turnout. 195 Cf. Christopher S. Elmendorf, Undue Burdens on Voter Participation: New Pressures for a Structural Theory of the Right to Vote?, 35 Hastings Const. L.Q. 643, 657 (2008) (arguing by focusing on turnout impacts in vote-denial cases arising under Constitution, judges can bridge normative differences regarding “reasonable voter” capacities).

3. Summary. — This Part has presented one account of how the core of section 2 could be implemented using evidentiary presumptions and current survey data, in keeping with the Shelby County Court’s sense that race-based remedial measures ought to be justified by the current risk of unconstitutional race discrimination. Ours will not be the final word. Some readers may disagree with our gloss on the core of section 2. Oth­ers may see different and perhaps better ways to craft the presumptions. What we hope to have shown is that it is at least feasible to answer—pre­sumptively—some of the recurring questions in section 2 cases using na­tional survey data, rather than the precinct-level vote tallies that have been the bread and butter of section 2 litigation so far.

We also hope to have persuaded the reader that these presumptions, if implemented with off-the-shelf statistical models, could enable section 2 to function more like section 5 in regions of the country where the pre­sumptions operate to shift evidentiary burdens to the defendants. Redis­tricters in such locales who do not provide minority communities with “roughly proportional” opportunities for representation would very likely face a section 2 lawsuit in which defendants would carry the burden of disproving central elements of the case. Vote-denial claims would also become easier for civil rights groups to litigate.

What remains to be established is that the courts have authority to es­tablish the presumptions and, further, that it is feasible to implement (most of) the presumptions using national survey data rather than case-specific studies carried out in particular defendant jurisdictions. 196 If the presumptions required original, case-specific research, they would be much more expensive to apply. We turn to these questions in the next Parts.

III. Authority, Legal and Otherwise

It is one thing to say that section 2 could be made to function like section 5 if Congress authorized an administrative agency to promulgate substantive rules about evidentiary presumptions. 197 Indeed, one possible legislative response to Shelby County would be to leave section 5 and the now-invalidated coverage formula as is, while authorizing DOJ to put new teeth into section 2 with substantive rules about evidentiary presumptions. It is quite another to maintain that section 2 can function similarly without any intervening action by Congress. This Part addresses two objections to our position: first, that the courts lack legal authority to establish the kinds of eviden­tiary presumptions suggested in Part II; second, that irrespective of legal authority, the courts cannot reasonably be expected to establish such pre­sumptions.

A. Legal Authority to Create the Presumptions

The proposition that courts lack legal authority to establish eviden­tiary presumptions under section 2 has little force. As one of us ex­plained in previous work, section 2 is best understood as a common law statute. 198 Elmendorf, Making Sense of Section 2, supra note 23, at 448–55. It delegates authority to the courts to implement loosely stated substantive and evidentiary norms. 199 See id. at 417–48 (discussing “helpful guidance” provided to courts by section 2 and legislative history in implementing equal voting norms). The legislative history makes clear that section 2’s results test was supposed to alleviate some of the eviden­tiary burdens associated with conventional intent tests in constitutional law, 200 See id. at 421–27 (noting legislative history reveals plaintiffs “may not be re­quired” to meet “conventional evidentiary standards”). but the courts were given broad discretion to shape the law going forward.

The courts have not shied from exercising this discretion. The statu­tory text instructs courts to base section 2 liability determinations on the “totality of circumstances,” but in the Supreme Court’s first vote dilution case under the results test, a four-Justice plurality tried to boil the matter down to whether a politically cohesive minority community had been consistently defeated at the polls. 201 See Thornburg v. Gingles, 478 U.S. 30, 63 (1986) (plurality opinion) (Brennan, J.) (“[M]ultimember districts may impair the ability of blacks to elect representatives of their choice where blacks vote sufficiently as a bloc as to be able to elect their preferred candidates in a black majority, single-member district and . . . a white majority votes . . . as a bloc . . . to defeat the candidates . . . .”). Some years later the Court resusci­tated the “totality of circumstances” inquiry and in doing so made cen­tral a factor that is not even mentioned in the legislative history: propor­tionality between the number of MODs and the minority’s population share. 202 See Johnson v. De Grandy, 512 U.S. 997, 1013–24 (1994) (finding totality of cir­cumstances did not support vote dilution claim because number of minority districts was “substantially proportional” to minority population share); Katz et al., supra note 30, at 730–31 (reporting of eighteen published cases in which courts made findings on pro­portionality, “[t]he 10 lawsuits that found proportionality identified no violation of Section 2,” and of the five lawsuits that “found a lack of proportionality[,] . . . four identi­fied a Section 2 violation”).

The lower courts have already developed rebuttable presumptions and burden-shifting rules in response to the Supreme Court’s signals. Thus, after Gingles, a number of courts held that a showing of minority political cohesion plus white bloc voting gives rise to a “strong presump­tion” of section 2 liability. 203 The presumption also requires that the minority community be large enough to satisfy the first Gingles factor. For cases recognizing this presumption, see supra note 111. Other courts, struggling with the question of whether white bloc voting is “legally significant” only if “caused” by the race of the candidate (or voters), held that a showing of racially polar­ized voting in biracial elections gives rise to a rebuttable presumption of race-discriminatory causation. 204 See supra note 111 (listing cases considering causation in totality of circum­stances inquiry). Courts have also used rules of thumb about the minority population share needed to ensure that an electoral district functions as an MOD. 205 See supra notes 182–183 and accompanying text (discussing “65%” rule of thumb for identifying minority opportunity districts). In short, presump­tions and burden-shifting rules are already embedded in the warp and woof of section 2.

Crafting burden-shifting rules and presumptions to implement broadly worded statutes is a familiar exercise for the courts. Judges put teeth into Title VII and other civil rights statutes with judge-made burden-shifting rules. 206 Plaintiffs’ showing of a racially disparate impact shifts the burden to the defen­d­ant to come forth with a legitimate rationale for the challenged law or practice, after which the plaintiff bears the ultimate burden of showing that the measures at issue are not reasonably necessary to serve the defendant’s legitimate interests. McDonnell Douglas Corp. v. Green, 411 U.S. 792, 802–04 (1973). Courts also borrowed evidentiary rules of thumb put forth in Equal Employment Opportunity Commission (EEOC) guide­lines, such as presuming a legally significant disparate impact where mi­norities are hired by an employer at less than four-fifths of the rate of white hiring. 207 See, e.g., Ricci v. DeStefano, 557 U.S. 557, 586–87 (2009) (finding prima facie case of disparate-impact liability shown, as EEOC’s 80% rule was violated); Watson v. Fort Worth Bank & Tr., 487 U.S. 977, 995 n.3 (1988) (noting lower courts’ use of EEOC’s four-fifths rule); Connecticut v. Teal, 457 U.S. 440, 443 n.4 (1982) (acknowledging district court’s finding of disparate impact under EEOC’s 80% rule). In antitrust law, the courts went further, deeming certain business arrangements per se anticompetitive. 208 See Andrew I. Gavil, Moving Beyond Caricature and Characterization: The Modern Rule of Reason in Practice, 85 S. Cal. L. Rev. 733, 744–45 (2012) (discussing shift from rule of reason analysis towards per se liability). The voting rights analogue would be a ban on literacy tests, which Congress applied to the covered jurisdictions in 1965 and extended nationally in 1970. Voting Rights Act Amendments of 1970, Pub. L. No. 91-285, 84 Stat. 315. Later courts relaxed some of the per se rules in light of new economic theory and evidence. 209 See Gavil, supra note 208, at 751–59 (noting courts’ recognition of inadequacy of per se liability); see also Andrew I. Gavil, William E. Kovacic & Jonathan B. Baker, Antitrust Law in Perspective: Cases, Concepts and Problems in Competition Policy 202–11 (2d ed. 2008) (describing evolution and application of presumptions shifting burden of production—and in some cases, burden of proof—to defendant, even as Court moved away from per se rules).

The rebuttable presumptions sketched in Part II are consistent with the notion that lawmaking by common law courts should be evolution­ary, not revolutionary. Each presumption serves to implement a norm that is already central to section 2 liability determinations. And because the pre­sumptions would be rebuttable, they are compatible with the statutory directive to base liability determinations on the totality of circumstances. 210 This Article imposes no limit on the “circumstances” that might be invoked to re­but inferences from the presumptions.

To the extent that our presumptions would work a large change in the law of section 2, the change is in the datasets and statistical tech­niques on which courts and litigants rely. 211 To be sure, the effect of this change in the law could be substantial in that section 2 claims would probably become fairly easy to win in some parts of the country, and quite difficult to win in other areas. But even this would only accentuate existing patterns. As Peyton McCrary has shown, the vast majority of successfully litigated or settled section 2 cases were brought in the formerly covered jurisdictions. Declaration of Dr. Peyton McCrary at 12, Shelby County v. Holder, 811 F. Supp. 2d 424 (D.D.C. 2011) (No. 1:10-cv-00651-JDB), http://moritzlaw.osu.edu/electionlaw/litigation/documents/Shelby-Dec1-1‌1-15-10.pdf [http://perma.cc/DN92-BCYN]. The empirical results presented in Part IV, infra, suggest that the same jurisdictions are likely to be the most vulnerable under the presumption-driven model for section 2 (at least with respect to claims of African Americans). National survey data on citi­zens’ political preferences and racial attitudes, analyzed using MRP, would partially supplant the current reliance on precinct-level election returns and ecological inference. 212 Using the proposed statistical model would only “partially” displace reliance on precinct-level results because in some cases, defendants may try to rebut the inference of racial polarization using data from local elections. But there is nothing in the text or legislative history of section 2 or in the Supreme Court’s authoritative constructions of the statute that compels judicial reliance on particular statistical tools. 213 Notably, the seminal decision in Thornburg v. Gingles, 478 U.S. 30 (1986) (plurality opinion) (Brennan, J.), which reoriented vote dilution law around the estima­tion of candidates’ vote shares by racial group, states that evidence that members of a ra­cial group tend to vote for the same candidates is simply “one way of proving the political cohesiveness necessary to a vote dilu­tion claim.” Id. at 56 (emphasis added).

Finally, on a purposive view of statutory interpretation, Shelby County’s negation of the section 5 preclearance regime counts strongly in favor of interpreting section 2 so that it works more like section 5, pro­vided of course that the reading does not push section 2 into the same constitutionally problematic territory. 214 See Cass R. Sunstein, Interpreting Statutes in the Regulatory State, 103 Harv. L. Rev. 405, 426–28 (1989) (defining purposive statutory interpretation). The VRA as enacted in 1965—and as reenacted in 1970, 1975, 1982, and 2006 215 Katz et al., supra note 30, at 646–47. —was predicated on a handful of core premises, which our model for section 2 adapts to the post–Shelby County world. The essential premises are, first, that the risk of unconstitutional race discrimination in the electoral process is higher in some parts of the country than in others and, second, that where this risk is high, mechanisms are needed to review, and in appropriate cases en­join, potentially discriminatory laws before they take effect, with the bur­den of proof borne by the alleged discriminator. 216 See supra notes 40–51 and accompanying text (describing key elements and goals of section 5).

The presumptions suggested here mesh these premises with section 2 while honoring Shelby County’s understanding of when it is permissible for legislation enforcing the Fourteenth and Fifteenth Amendments to single out states for special burdens. Shelby County faulted the preclear­ance coverage formula for distinguishing among states on the basis of old data that bore no apparent relationship to the current risk of unconstitu­tional race discrimination. 217 Shelby County v. Holder, 133 S. Ct. 2612, 2625–30 (2013) (“[T]he coverage for­mula that Congress reauthorized in 2006 ignores these developments, keeping the fo­cus on decades-old data relevant to decades-old problems, rather than current data reflect­ing current needs.”). The presumptions would be implemented using current data and would maintain a close connection between the risk of section 2 liability (higher where evidentiary burdens are shifted to the defendant) and the risk of unconstitutional state action.

B. Judicial Competence

That judges have legal authority to implement our approach does not mean they will be able to do it or do it effectively on their own. Sec­tion 2 cases present very difficult technical and legal questions. Our sense from reading published section 2 opinions is that the first priority for many judges is simply to avoid embarrassment.

Rather than wrestle with the reliability of different techniques of eco­logical inference, judges continue to accept questionable methods on the basis of an offhand, decades-old footnote from the Supreme Court characterizing the methods as “standard in the literature.” 218 Gingles, 478 U.S. at 53 n.20 (plurality opinion) (Brennan, J.). Some judges do seem willing to accept the new, bet­ter methods only if their results accord with the older methods. D. James Greiner, The Quantitative Empirics of Redistricting Litigation: Knowledge, Threats to Knowledge, and the Need for Less Districting, 29 Yale L. & Pol’y Rev. 527, 532–33 (2011) [hereinafter Greiner, Quantitative Empirics]. Harvard statistician and law professor Jim Greiner wrote several outstanding pa­pers critiquing standard ecological inference techniques and offering better alternatives; 219 Greiner, Ecological Inference, supra note 78; Greiner & Quinn, Exit Polling and Racial Bloc Voting, supra note 79; Greiner, Quantitative Empirics, supra note 218; Greiner, Re-Solidifying Racial Bloc Voting, supra note 69; D. James Greiner & Kevin M. Quinn, R×C Ecological Inference: Bounds, Correlations, Flexibility and Transparency of Assumptions, 172 J. Royal Stat. Soc’y: Series A 67 (2009) [hereinafter Greiner & Quinn, R×C Ecological Inference]. his work has left no impression on the courts. 220 A Westlaw search turned up only one opinion that cites Greiner’s work on eco­logi­cal inference. Levy v. Lexington County, No. 3:03-3093-MBS, 2012 WL 1229511, at *6 (D.S.C. Apr. 12, 2012) (discussing Greiner, Re-Solidifying Racial Bloc Voting, supra note 69). In Levy, a wise judge appointed Greiner’s collaborator Kevin Quinn to advise the court on statistical methods. Id. at *4 n.4. The continued acceptance of statistical techniques that are “standard” per their use in prior cases (even if unreliable) saves the judge from po­tential embarrassment, for if she errs, she makes only the same mistake as her peers.

As for the law, section 2 offers an easy out to judges who don’t want to venture a transparent interpretation of the statute’s substantive and evidentiary norms: Glide past the conceptual questions, duly note that the statutory text calls for liability determinations to be based on the “to­tality of circumstances,” and then recite a long list of circumstances that nom­inally ground your decision.

The project of crafting rebuttable presumptions to implement section 2 requires judges to take some risks—especially if plaintiffs are invited to make evidentiary showings based on new-fangled statistical tech­niques. And to fully realize the promise of our presumption-driven approach to section 2, many judges must take the plunge, define the pre­sumptions similarly, and agree on the datasets and models that litigants may use to determine which presumptions apply in a given case.

Likelihood-of-success determinations will become straightforward only if the courts coordinate on a common model and dataset, as well as common definitions of the presumptions. The incremental, disaggre­gated process of common law adjudication makes this coordination diffi­cult, particularly given the wide range of plausible evidentiary presump­tions. 221 See supra Part II (discussing wide range of possible presumptions to make section 2 viable stand in for section 5). The obvious alternative is to assign responsibility for developing the presumptions to an administrative agency. Agencies may compel ju­dicial coordination by issuing rules with the force of law, agencies have the necessary technical expertise, and agencies can involve a much broader swath of the public in developing the law via advisory commit­tees and notice-and-comment rulemaking. But section 2 does not dele­gate rulemaking authority to any agency. DOJ litigates section 2 cases from time to time but has never issued enforcement guidelines or inter­pretive rules under section 2. 222 Interview with Michael Pitts, Professor, Ind. Univ. Robert H. McKinney Sch. of Law, and former trial attorney, Voting Section, U.S. Dep’t of Justice.

It is conceivable that DOJ could nonetheless induce judicial coordi­nation by issuing interpretive rules under section 2. These rules would be advisory only, not binding, but they would be owed Skidmore deference. 223 Under Skidmore v. Swift, 323 U.S. 134 (1944), judicial deference to agency posi­tions varies according to the quality of the agency’s decisionmaking process, taking ac­count of any “factors which give [an interpretation] power to persuade, if lacking power to control.” Id. at 139–40. We agree with Professor Nou that, in the election administration context, the Skidmore framework counsels for calibrating deference to “the institutional role of the actors authoring the interpretive documents and, specifically, the degree to which they are internally politically insulated.” Jennifer Nou, Sub-Regulating Elections, 2013 Sup. Ct. Rev. 135, 152. Title VII provides an instructive analogy: Interpretive guidelines issued by the EEOC have not been treated as binding on the courts, but they are given some weight and have played an important role in fleshing out Title VII’s disparate-impact standard. 224 For example, the “four-fifths rule” in disparate impact cases originated with EEOC guidelines. See cases cited supra note 207. To be sure, the track record of judicial deference to EEOC positions is checkered. See Melissa Hart, Skepticism and Expertise: The Supreme Court and the EEOC, 74 Fordham L. Rev. 1937, 1949–61 (2006) (attrib­uting skepticism of some judges to perception EEOC is or has been pursuing narrowly po­litical agenda and not basing its rules on any material or technical expertise). As discussed next, these same concerns could well arise with DOJ-issued guidelines under section 2, but the concerns can be blunted if the agency takes them into account ex ante. To be sure, some considerations cut against judicial deference, such as the fact that DOJ has often been accused of partisanship in the administration of voting rights laws. 225 In the 2000s, under President George W. Bush, many career staffers in the voting section departed and were replaced with new hires whom critics on the left derided as partisan hacks. Charlie Savage, Report Examines Civil Rights During Bush Years, N.Y. Times (Dec. 2, 2009), http://www.nytimes.com/2009/12/03/us/politics/03rights.html?_‌r=1 (on file with the Columbia Law Review). Under President Obama, DOJ has taken a strong stance against voter-ID laws and other measures favored by Republicans, leading Republicans to levy the charge of unlawful partisanship. See Roger Clegg & Hans A. von Spakovsky, More Voting-Rights Challenges from Holder, Nat’l Rev. Online (July 29, 2013), http://www.nationalreview.com/article/354618/more-voting-rights-challenges-holder-roger-clegg-hans-von-spakovsky [http://perma.cc/7SFG-254H] (“[The Obama] administration has a highly politicized Civil Rights Division, and using the Voting Rights Act to achieve partisan ends is nothing new for it.”); Hans A. von Spakovsky, Every Single One: The Politicized Hiring of Eric Holder’s Voting Section, PJ Media (Aug. 8, 2011), http://pjmedia.com/blog/every-single-one-the-politicized-hiring-of-eric-holder’s-voting-section/ [http://perma.cc/4RD5-G28C] (“[T]he Civil Rights Division at the Department of Justice is engaging in politicized hiring in the career civil service ranks . . . nearly unprecedented in scope and significantly eclips[ing] anything the Bush administration was even accused of doing.”).
Inspector General reports in 2008 and 2012 gave credence to some of the accusations of partisanship. U.S. Dep’t of Justice, Office of the Inspector Gen., A Review of the Operations of the Voting Section of the Civil Rights Division 117–80 (2013) (addressing allegations of Voting Section employees mistreated on basis of political ideology); U.S. Dep’t of Justice, Office of the Inspector Gen. & U.S. Dep’t of Justice, Office of Prof’l Responsibility, An Investigation of Allegations of Politicized Hiring and Other Improper Personnel Actions in the Civil Rights Division (2008) (summarizing investigation “into allegations that political or ideological affiliations were considered in hiring, transferring, and assigning cases to career attorneys in the Civil Rights Division”); U.S. Dep’t of Justice, Office of the Inspector Gen. & U.S. Dep’t of Justice, Office of Prof’l Responsibility, An Investigation of Allegations of Politicized Hiring in the Department of Justice Honors Program and Summer Law Intern Program (2008) (describing results of investigation into “whether the political or ideological affiliations of applicants were improperly considered in the selection of candidates for the [Attorney General]’s Honors Program and the Summer Law Intern Program . . . from 2002 to 2006”); U.S. Dep’t of Justice, Office of Prof’l Responsibility & U.S. Dep’t of Justice, Office of the Inspector Gen., An Investigation of Allegations of Politicized Hiring by Monica Goodling and Other Staff in the Office of the Attorney General (2008) (investigating “allegation that Goodling,” DOJ’s White House Liaison, “inappropriately used political or ideological affiliations in the hiring process for career [DOJ] employees”). But if DOJ structured the process for developing the rules so as to preempt the charge of partisanship, 226 For example, by assigning rule development to a technical or bipartisan body not controlled by political appointees at DOJ. appellate courts might well defer because DOJ rulemaking would save the courts from having to make difficult tech­nical determinations and because adoption of the rules would make section 2 litigation less subject to the whims of individual trial judges. 227 See Christopher S. Elmendorf, Advisory Rulemaking and the Future of the Voting Rights Act, 14 Election L.J. 260, 270–75 (2015) (assessing costs and benefits of ju­dicial deference to DOJ guidelines under section 2 from perspective of appellate judges). Professor Ross argues that the Supreme Court has shown no inclination to defer to agen­cies on questions that implicate constitutional interpretation. Bertrall Ross, Denying Deference: Civil Rights and Judicial Resistance to Administrative Constitutionalism, 2014 U. Chi. Legal Forum 223, 228. Ross’s argument implies—and we agree—that the Court would not defer to DOJ’s decisions about, for example, whether section 2 still provides a remedy for constitutional violations that occurred during the Jim Crow era. But so long as DOJ works from the Court’s own constitutional premises, Ross’s argument provides no basis for doubting judicial deference to agency positions on technical questions about evidentiary presumptions. As explained supra notes 94–97, our presumptions build on an interpretation of section 2 meant to reconcile the results test with the current Supreme Court’s constitutional understandings.

If judges generally followed DOJ’s recommendations, this would greatly reduce uncertainty about how the presumptions cut in a given case. Rather than reinventing the wheel, plaintiffs’ experts could simply download the gold-standard model and dataset from DOJ’s website, and run it for the racial group(s) and jurisdiction at issue in the case. Once the model has been accepted by a few courts, it will no longer be worth­while for defendants to attack it in ordinary cases. 228 To be sure, in rare cases where the political stakes are very high, it may be worth­while for defendants to attack the model, just as it was worthwhile in the pre–Shelby County era for some covered jurisdictions to seek preclearance from the District Court of the District of Columbia rather than DOJ in certain high-stakes, politically charged cases. See, e.g., Rick Hasen, Alabama Bypassing DOJ in Favor of Court Seeking Preclearance of Redistricting Plan, Election Law Blog (Sept. 14, 2011, 7:54 am), http://electionlawblog.org/
?p=23078 [http://perma.cc/R4TK-GPSA] (“The pattern continues of Republicans in cov­ered jurisdictions not trusting the Obama Administration’s DOJ on Voting Rights Act pre­clearance of controversial redistricting plans.”). At this point, a legal regime that formally requires plaintiffs to make evidentiary showings with respect to constitutional risk and disparate impact would function as if it were a regime in which certain geographically delimited jurisdictions were formally presumed to be at risk of unconstitutional race discrimina­tion in the electoral process and in which certain electoral structures (in these jurisdictions) are presumed to have a disparate impact. 229 We recognize that the gold-standard model would likely evolve over time, in keep­ing with advances in political science and statistics. Turnover in DOJ’s leadership may also lead to a revisiting of the guidelines. Continued judicial acceptance of the guidelines would of course depend on the credibility/impartiality of the process by which the Department updates them. The gap between section 2 and now-defunct section 5 would be much diminished.

But what if DOJ is rebuffed by the courts or stays on the sidelines? That would retard the process of developing a presumption-driven section 2 implemented with national survey data, but gradual change would remain possible. For example, rather than asking courts to adopt new presumptions under Gingles, litigants could introduce survey evi­dence of racial polarization and racial stereotyping to supplement vote-share evidence under Gingles, or as part of the “totality of circumstances.” Once one court gives some weight to this evidence, other litigants will have incentives to bring it forward in the next case, and judges will have to weigh the advantages and disadvantages of conventional and survey-based evidence. Over time, even the most cautious, incrementalist judges are likely to give progressively more weight to survey data because survey-based estimates do not suffer from the problems that make it difficult to infer preference polarization and voter discrimination from vote shares for “minority candidates of choice,” 230 See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 50–71) (explaining perils of trying to infer political cohesion and polarization from votes in actual elections). and because survey-based estima­tion does not involve ecological inference. 231 See id. (manuscript at 71–73, 81–89) (explaining ecological inference depends on racial homogeneity assumptions similar to those Supreme Court has disavowed); see also Greiner, Re-Solidifying Racial Bloc Voting, supra note 69, at 465–68 (explaining eco­logical inference often yields unreliable estimates where there are more than two racial groups and significant residential integration).

As evidence from national surveys starts to play a larger role in the ad­judication of section 2 cases, the decisions themselves will provide in­creasing guidance about how pending or prospective cases are likely to be resolved. For example, if the level of racial polarization in jurisdiction A is deemed “legally significant,” and if the same or higher levels of po­larization exist in jurisdiction B per the data sources and statistical models used in the previous case, then lawyers for both parties in a newly filed case against jurisdiction B should have a pretty good sense of whether a court is likely to find legally significant polarization in B.

One might suppose that this would be true irrespective of whether the second case is litigated primarily on the basis of local election data or national survey data. Not so. If the case in B depends on local election returns, then the plaintiff will have to pay an expert to retrieve local elec­tion files from county courthouses, to digitize those records, to estimate the correlation between race and vote choice in each election, and then to make an argument about which elections are most probative of racial polarization in the community. That argument would turn on factors such as the race of the candidates, their backing from local political elites within the minority and white communities, incumbency, the responsibil­ities of the office in question, the date of the election, and any other “spe­cial circumstances” that arguably bear on the degree to which racial polarization in vote choice does or does not signify racial polarization in enduring political preferences. 232 See generally Elmendorf et al., Racially Polarized Voting, supra note 60 (manu­script at 33) (“[I]nterracial elections are more probative than same-race elections [and] endogenous elections are more probative than exogenous elections.”); Katz et al., supra note 30, at 668–70 (discussing factors relevant to identifying probative elections). Previous cases will provide some guid­ance about the factors to consider in judging probativeness, 233 See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 24–27, 29–31) (discussing different courts’ interpretations of probativeness). but they cannot resolve the ultimate question of how much weight to assign to each election introduced in the case against B. 234 This is so because the weight assigned to each election depends on the proba­tiveness of every other election in the record. (It is telling that the lower courts have consistently rejected arguments for numeric bloc-voting “cutoffs” under Gingles. 235 See Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at 33) (“As things stand today, there are no established quantitative cutoffs to distinguish polar­ized from non-polarized communities.”). ) By contrast, if the decision in case A turned on a measure of ideology or racial attitudes derived from national sur­veys 236 See infra notes 242–250 and accompanying text (describing use of national sur­vey data to develop reliable measures of local racial attitudes). and if the same survey data and statistical models are deployed in case B, the holding in A will be very instructive about the likelihood of liability in case B. And the cost of figuring out how the holding in A cuts in case B will be minimal, assuming that the pertinent dataset and statis­tical models are in the public domain. 237 Large-sample surveys by political scientists are conventionally put into the public domain within a year or two of their completion. Standard tools for estimating local opin­ion from national surveys are also in the public domain. E.g., Mike Malecki, Multilevel Regression and Poststratification, Github (May 1, 2014), https://github.com/malecki/‌mrp [https://perma.cc/S8MY-JEKE] (providing public domain package in statistical program­ming language R for MRP).

Finally, it is worth noting that courts can and often do create very in­formative evidentiary guideposts without using the label “presumption.” For example, though courts have rejected the proposition that propor­tionality between minority population share and the number of majority-minority districts is an absolute defense to liability in vote dilution cases, 238 See Johnson v. De Grandy, 512 U.S. 997, 1013–14 (1994) (explaining different lev­els of proportionality may still violate VRA). courts have nonetheless signaled that proportionality is very im­portant 239 See, e.g., Black Political Task Force v. Galvin, 300 F. Supp. 2d 291, 311 (D. Mass. 2004) (“One of the most revealing questions a court can ask in assessing the totality of the circumstances is whether the affected districts exhibit proportionality . . . .”); Campuzano v. Ill. State Bd. of Elections, 200 F. Supp. 2d 905, 908 (N.D. Ill. 2002) (“For a plan to pro­vide minority voters equal participation in the political process, it must generally provide a number of ‘effective’ majority-minority districts that are substantially proportionate to the minority’s share of the state’s population.”). and practitioners have had no trouble reading the signal. 240 See, e.g., Hebert et al., Realist’s Guide, supra note 30, at 36 (observing “‘propor­tionality,’ or lack thereof,” is “particularly important” factor in vote dilution cases). Similar conventions may well emerge regarding survey-based evidence of constitutional risk, racial polarization, and minority opportunity, even without formal recognition of the conventions as evidentiary presumptions.

In summary, the courts quite clearly have legal authority to create ev­identiary presumptions along the lines suggested in Part II. The incre­mental process of common law adjudication is not ideal for this purpose, but even a gradual increase in judicial reliance on national survey data in section 2 cases would yield a corresponding increase in the degree to which section 2 precedents are informative about the likely resolution of future cases (in keeping with the model of a presumption-driven section 2). DOJ may be able to accelerate this process by issuing advisory guide­lines under section 2.

What remains to be considered is where a presumption-driven section 2 would likely have the most bite. In particular, would the regions of the country formerly subject to the section 5 preclearance regime end up “covered” de facto by evidentiary presumptions under section 2? This is the subject of the next Part.

IV. Model Building and Results

This Part introduces the statistical machinery for generating esti­mates of public opinion within small geographic units from national sur­veys and presents some initial empirical results. We consider our results provisional because they are based on models that, while facially reason­able, have not been validated with out-of-sample data, and because the corresponding presumptions could be defined in other ways. 241 See supra note 122 and accompanying text (describing other methods of defin­ing presumptions).

A. Tools for Estimating Racial-Group Opinion Within Subnational Geographic Units 242 Some of the description of methodology in this section also appears in Elmendorf & Spencer, Preclearance, supra note 19, at 1156–57.

Given a national survey dataset, there are two commonly used tech­niques for estimating opinion within particular racial groups in discrete geographic units. One is to disaggregate the data by race and geography, and, if the survey is not an equal-probability sample of the population of interest, to reweight the disaggregated data so that it matches known de­mographics of the target population. 243 Because of sampling challenges and non-response (and sometimes by design), no survey is a true equal-probability sample of the target population. On reweighting, see generally Carl-Erik Särndal & Sixten Lundström, Estimation in Surveys with Nonresponse 43–53 (2005). The second approach is to model responses to the survey question as a function of the respondent’s demo­graphic and geographic attributes, making inferences about public opin­ion in one geographic unit based on the responses of similarly situ­ated respondents elsewhere. This is done using MRP.

To illustrate the difference between the two approaches, imagine that we want to estimate the racial attitude of white people in the city of Boston and in the state of Massachusetts. Assume we have data from a nationally representative sample of 2,000 white citizens. Because Boston contains about 0.2% of the national population and Massachusetts about 2.1%, our sample would include, on average, only about four Bostonians and forty-two respondents from Massachusetts. With sample sizes of just a handful or even a few dozen respondents, random selection will quite often yield survey samples that are considerably more or less prejudiced than the actual population of white Bostonians or Bay Staters. 244 In technical terms, the standard deviation of a sample mean or proportion is in­versely related to the square root of the sample size. Large representative survey samples of the adult U.S. population are rare. One important sample is the Cooperative Congressional Election Survey (CCES), which was created for the express purpose of stud­ying voter opin­ion within small geographic units, and by sample size is the largest regularly conducted survey of American voters, with about 100 respondents per congressional dis­trict. See generally Stephen Ansolabehere & Douglas Rivers, Cooperative Survey Research, 16 Ann. Rev. Pol. Sci. 307 (2013) (describing CCES). Perhaps the four Bostonians who took the survey are young, highly educated women who (let us assume) tend to be much less prejudiced than older, less affluent men. If so, the average level of prejudice in the survey sam­ple of Bostonians is likely to badly understate average prejudice in the population, at least if there are a lot of old, poorly educated men in Bos­ton. Reweighting the data to the known distribution of “demographic types” in the population is not a good fix, because some types are likely to be absent from the sample entirely (only four Bostonians took the sur­vey), and because the small number who did take the survey may be highly unrepresentative of their type (not all young, highly educated women have the same opinions). 245 Also, if some units in the sample receive much more weight than other units, this blows up the variance of the sample-mean estimator—meaning that the estimate is a very imprecise proxy for the true population mean. See Roderick J.A. Little & Donald B. Rubin, Statistical Analysis with Missing Data 49 (2d ed. 2002) (“[Propensity weighting] removes nonresponse bias, but . . . may yield estimators with extremely high variance be­cause re­spondents with very low estimated response propensity receive large nonresponse weights and may be unduly influential in estimates of means and totals.”).

One solution to this problem is to pool together the responses from multiple surveys over time and then disaggregate the pooled surveys by race and geography. This strategy holds some promise, but for our pur­pose requires proprietary data that have not yet been released for public use. 246 This footnote explains the problem. We need a summary measure of policy agreement/disagreement between racial communities. The best summary measure at this time is an “ideal point” scaled from policy preferences. See infra notes 275–280 and accompanying text (discussing advantages of scaled ideal point measure, compared to self-reported ideology or partisanship). But most national surveys do not include the same policy questions, so ideal points scaled from the policy questions on each survey are not comparable across surveys. Political scientists Chris Tausanovitch and Chris Warshaw re­cently solved this problem by including “bridging” questions on several modules on the CCES; this enabled them to create a “superset” of 275,000 respondents with ideal points on the same scale. See Chris Tausanovitch & Christopher Warshaw, Measuring Constituent Policy Preferences in Congress, State Legislatures, and Cities, 75 J. Pol. 330, 332 (2013) (describing compilation of surveys in creation of “supersurvey”). After this dataset is re­leased into the public domain, it may be feasible to create reasonably precise estimates of the distribution of racial group opinion within congressional districts, counties, and other small geographic units by disaggregation (after reweighting the data to match local demographics). Another possible solution would be to conduct original surveys within the unit of interest. This strategy cuts against our goal of making (presumptive) section 2 liability easy to ascertain, 247 The problem is not simply one of cost. If original surveys must be conducted for each case, then potential defendants may not be able to anticipate liability ex ante (with­out conducting surveys themselves), and there are likely to be case-specific disputes about the survey methodology, the qualifications of the experts who conducted the survey, etc. so we do not pursue it further here. 248 If we were only interested in vote shares, there would be a fourth option, ecologi­cal inference, but for reasons explained earlier we hope to avoid it. See supra text accom­panying notes 78–82 (noting conclusions drawn from ecological inference are “tenuous” when more than two racial groups are involved or if jurisdiction is significantly integrated).

Given a single national survey, MRP improves on disaggregation by us­ing demographic and geographic identifiers to make inferences about similarly situated respondents. From publicly available Census data, we know how many Bostonians are young, highly educated women and how many are old, poorly educated men. 249 For example, Boston is 48% male, 17% of all residents are under eighteen years old, and 10% are older than sixty-five. The city is 54% white and 24% black and 36% speak a lan­guage other than English in the home. Just 20% of Boston residents graduated from col­lege (compared to the national average of 26.3%). The median household income in Boston is $53,601 with 21% living below the poverty line. United States Census Bureau, American FactFinder, http://factfinder.census.gov/faces/nav/jsf/pages/index.xhtml (enter “Boston, MA” in search bar; then follow “General Population and Housing Characteristics” hyperlink) (last visited Oct. 14, 2015) (data on file with the Columbia Law Review) (provid­ing 2010 Census data). In both 2008 and 2012, about 78.5% of the city’s voters voted for Obama in the general election. Sec’y of the Commonwealth of Mass., 2008 President General Election, Suffolk County, http://electionstats.state.ma.us/elections/view/14274/filter_by_county:Suffolk [http://perma.cc/2V7G-7RU3]; Sec’y of the Commonwealth of Mass., 2012 President General Election, Suffolk County, http://electionstats.state.ma.us/elections/view/22515/filter_by_county:Suffolk [http://perma.cc/U2DU-3NB8]. To the extent that age, education, and sex predict racial attitudes, or any other opinion of interest, we can use what we learn from the 2,000-person national survey about old, poorly educated men (and other demographic types) elsewhere in the country to estimate Bostonians’ opinions, weighting the estimates by what the Census tells us is the frequency of each demographic combina­tion in Boston.

To be sure, the racial attitudes of white people in Boston or Massachusetts may diverge from those of their demographic doppel­gang­ers elsewhere. If so, imputing to Bostonians the average attitude of their respective demographic types could generate estimates of citywide opin­ion that are badly off base. MRP accounts for this by allowing the inter­cept term in the regression model to vary with geography. In other words, the estimated racial attitude of white Bostonians is adjusted up­ward or down­ward depending on whether the survey respondents in Boston are more or less prejudiced than their demographic counterparts elsewhere. The size of the adjustment is proportional to the amount of information in the sample about people from the geographic unit. If lots of Bostonians answered the survey, we would have more confidence that any difference between their answers and those of their demographic counterparts reflects actual differences between the underlying popula­tions. In this case, the statistical model would make a corresponding large adjustment. However, if few Bostonians were surveyed, the adjust­ment would be small. 250 See Andrew Gelman & Jennifer Hill, Data Analysis Using Regression and Multilevel/Hierarchical Models 6–8, 252–54 (2007) (explaining multilevel regression pro­duces precision-weighted average of “complete pooling” and “no pooling” estimators). In the running example, the complete-pooling estimator is an average of responses from all geographic units; the no-pooling estimator is an average of just those responses from the target unit, i.e., Boston. The no-pooling estimator would be imprecise (hence given little weight) if there were few responses from the target unit.

The flexibility of MRP extends beyond weighted adjustments based on sample size. With just four or even forty-two respondents in each geo­graphic unit, it would be difficult to infer very much about geographic variation in racial attitudes. But just as we can “borrow” information from old, white men who live elsewhere to improve our estimates of the atti­tudes of Bostonians, so too can we borrow information from other geo­graphic units to better understand how the attitudes of people in Boston are likely to differ from the attitudes of demographically similar people elsewhere. We can, for example, model the intercept term for small geo­graphic units as a function of some larger unit in which it is nested—Boston as a function of Massachusetts or the Northeast Region. The larger unit, of course, includes more survey respondents. This modeling decision allows evidence of deviations between (1) the racial attitudes of particular demographic types within the midsized geographic units (Northeast Region) and (2) the attitudes of the same demographic types in the na­tional sample, to serve as evidence of the difference be­tween people in the small unit (Boston) and citizens across the nation. Another option is to model the intercept for geographic units as a func­tion of unit-level characteristics, such as residential segregation, that the­ory suggests are correlated with individual attitudes. 251 See id. at 254–59 (discussing models featuring unit-level predictors). Still another possibility is to build in interactions between individual-level predictors and attributes of the geographic unit. For example, the correlation be­tween residential inte­gration and white stereotyping of blacks may de­pend on the income of white respondents. 252 See Wendy K. Tam Cho & Neil Baer, Environmental Determinants of Racial Attitudes Redux: The Critical Decisions Related to Operationalizing Context, 39 Am. Pol. Res. 414, 416 (2011) (“[S]cholars have found that socioeconomic conditions and eco­nomic disparities mediate how various social contexts translate into prejudicial atti­tudes.”). The “racial threat hypothesis” posits that whites who live near large black or other minority populations will feel more threatened by the minority group and that this sense of threat will manifest itself in negative racial stereotypes. See Marisa Abrajano & Zoltan Hajnal, White Backlash 117–18 (2015) (suggesting “larger out-groups can repre­sent a threat to members of the in-group” because “proximity tends to enhance real or perceived competition for scarce resources”); V.O. Key, Southern Politics in State and Nation 665–68 (1949) (focusing attention on Southern counties with large black popula­tion, characterized as “hard core of the political South”); Lawrence Bobo & Vincent L. Hutchings, Perceptions of Racial Group Competition: Extending Blumer’s Theory of Group Position to a Multiracial Context, 61 Am. Soc. Rev. 951, 953 (1996) (noting indi­viduals who “face racially changing neighborhoods . . . are most likely to feel threatened by competition from members of other minority groups”); Claudine Gay, Seeing Difference: The Effect of Economic Disparity on Black Attitudes Toward Latinos, 50 Am. J. Pol. Sci. 982, 995 (2006) (addressing “behavior of white Americans, whose hostility toward minority outgroups rises in direct proportion to the size of the proximate minority popu­lation”). The racial threat hypothesis has been used to explain geographic variation in white support for policies such as affirmative action. See David Austen-Smith & Michael Wallerstein, Redistribution and Affirmative Action, 90 J. Pub. Econ. 1789, 1791 (2006) (“A low-income white voter . . . may prefer a party that opposes redistribution . . . .”); Caroline Tolbert & John A. Grummel, Revisiting the Racial Threat Hypothesis: White Voter Support for California’s Proposition 209, 3 St. Pol. & Pol’y Q. 183, 197 (2003) (“[O]ur findings that whites oppose affirmative action as their neighborhoods become more ra­cially diverse regardless of race is more consistent with the cultural backlash hypothe­sis . . . .”). But see Andrea Louise Campbell et al., “Racial Threat”, Partisan Climate, and Direct Democracy: Contextual Effects in Three California Initiatives, 28 Pol. Behav. 129, 141 (2006) (“The proportion of black residents in a county . . . did not affect vote choice on the affirmative action proposition.”). This hypothesis has also been used to understand geographic varia­tion in support for redistribution. See Alberto Alesino & Edward Glaeser, Fighting Poverty in the United States and Europe: A World of Difference 148–50 (2004) (analyzing effect of “racial fractionalization” on redistribution support); Austen-Smith & Wallerstein, supra, at 1790 (addressing effect of “social cleavages” on redistribution); Hersh & Nall, supra note 161, at 13 (addressing “‘spatial regime’” of state and substate regions in explaining voting behavior). Finally, the hypothesis is also helpful to under­standing geographic variation in support for harsh criminal laws and sanctions. See David Jacobs et al., Vigilantism, Current Racial Threat, and Death Sentences, 70 Am. Soc. Rev. 656, 660 (2005) (“Theory suggests that whites will make greater demands for punitive measures after expansions in black presence.”). If the racial threat hypothesis is correct, then black population size in the geographic unit should be negatively correlated with whites’ stereo­types, holding constant the demographic attributes of the white population. Conditional on black population size, it might also be the case that racial attitudes correlated with the degree of residential inte­gration, either because racially tolerant people are drawn to inte­grated neighborhoods or because quotidian interracial contact increases tolerance. This possibility can be accom­modated by modeling the intercept term as a function of the level of residential integra­tion in the geographic unit.

Once region- and unit-level characteristics such as black population size or residential integration have been incorporated into the model, intercepts can be estimated for small geographic units even if there are few or no respondents in a given unit. Intercept estimates will continue to reflect idiosyncratic information about respondents from particular units, for example the difference between the racial attitudes of Bostonians in the sample and the racial attitudes that the model predicts for them based on their demographics and geographic unit. The weight attached to idiosyncratic information is proportional to the number of respond­ents from the unit. 253 This means that models of opinion at the state level will be more strongly an­chored to the actual survey responses of people in the geographic unit (state) than models of opinion at the county level where there is more “shrinkage” to the average response of similarly situated respondents. See Gelman & Hill, supra note 250, at 254 (noting averages from counties with larger sample sizes yield multilevel estimates “close to the county aver­ages,” while averages from counties with smaller sample sizes are pulled closer to overall state averages).

The final step in the MRP modeling process, called poststratifica­tion, weights the estimated opinion of each demographic type (e.g., white men over the age of sixty-five who did not attend college), within a unit by the type’s share of the unit’s adult population. This yields an ap­proximation of the empirical distribution of opinion within the unit.

MRP has become very popular among political scientists as a way to es­timate public opinion on political issues within subnational jurisdic­tions. 254 See Gelman & Hill, supra note 250, at 8 (noting multilevel modeling is essential for forecasting state-by-state outcomes of U.S. presidential elections); Yair Ghitza & Andrew Gelman, Deep Interactions with MRP: Election Turnout and Voting Patterns Among Small Electoral Subgroups, 57 Am. J. Pol. Sci. 762, 763–64 (2013) (explaining statistical method and role of poststratification in modeling voting patterns); Jeffrey R. Lax & Justin H. Phillips, Gay Rights in the States: Public Opinion and Policy Responsiveness, 103 Am. Pol. Sci. Rev. 367, 369 (2009) (using MRP for “estimates of state-level policy-specific opinion”); Jeffrey R. Lax & Justin H. Phillips, How Should We Estimate Public Opinion in the States?, 53 Am. J. Pol. Sci. 107, 109 (2009) [hereinafter Lax & Phillips, How Should We Estimate?] (lauding simplicity of MRP analysis); Julianna Pacheco, Using National Surveys to Measure Dynamic U.S. State Public Opinion: A Guideline for Scholars and an Application, 11 St. Pol. & Pol’y Q. 415, 420 (2011) (describing MRP as “superior” aggrega­tion method); David K. Park, Andrew Gelman & Joseph Bafumi, Bayesian Multilevel Estimation with Poststratification: State-Level Estimates from National Polls, 12 Pol. Analysis 375, 377 (2004) (describing application of multilevel estimation with poststratifi­cation to elections). It has been used to estimate opinion within states, 255 See, e.g., Ghitza & Gelman, supra note 254, at 774–75 (using MRP to analyze vote choice and turnout of demographic subgroups within states); Lax & Phillips, How Should We Estimate?, supra note 254, at 120 (concluding MRP, “if implemented using a single, large national survey, produces estimates of state-level public opinion that are vir­tually as accurate as those it generates using 10 or more surveys”); Pacheco, supra note 254, at 419 (explaining “MRP . . . allows for the inclusion of various demographic predic­tors to estimate state public opinion”). For a user-oriented introduction to the methods, see Gelman & Hill, supra note 250, at 1–11 (providing brief introduction to multilevel regression modeling and motivations). congres­sional districts, 256 See Christoper Warshaw & Jonathan Rodden, How Should We Measure District-Level Public Opinion on Individual Issues?, 74 J. Pol. 203, 216 (2012) (“MRP yields esti­mates of issue-specific district public opinion that are consistently superior to disaggre­gated means or presidential vote shares.”). state legislative districts, 257 See Tausanovitch & Warshaw, supra note 246 (stating MRP estimates had higher correlation with 2008 presidential vote shares than disaggregated estimates at state legisla­tive district level); Warshaw & Rodden, supra note 256, at 218 (concluding MRP “can pro­duce reliable estimates for state legislative districts on many issues”); David E. Broockman & Christopher Skovron, What Politicians Believe About Their Constituents: Asymmetric Misperceptions and Prospects for Constituency Control 13–14 (July 17, 2015) (un­published manuscript), http://stanford.edu/~dbroock/papers/broockman_skovron_‌asymmetric_misperceptions.pdf [http://perma.cc/L5QT-XH3A] (describing steps taken to apply MRP to state legislative districts). cities, 258 See Tausanovitch & Warshaw, supra note 246, at 335–36 (stating MRP estimates had higher correlation with 2008 presidential vote shares than disaggregated estimates at city level). and even local school board districts. 259 Michael Berkman & Eric Plutzer, Evolution, Creationism, and the Battle to Control America’s Classrooms 83–87 (2010) (detailing MRP’s advantages in estimating local support for teaching of evolution). The technique has been validated—and shown to out­perform disaggregation—in a number of instances where mean public opinion (or something close to it) within small geographic units can actually be observed, such as vote shares for presidential candidates, or vote shares on local referenda that address an issue that also appears on a national survey. 260 See, e.g., Warshaw & Rodden, supra note 256, at 211 (“[T]he MRP estimates are better predictors of referendum results than the disaggregated estimates. The MRP esti­mates have higher correlations with the referenda results for each issue and state, and generally smaller mean absolute errors.”).

But for legal applications, it is also important to appreciate MRP’s lim­itations. MRP is an example of what statisticians call parametric or model-based estimation techniques. MRP estimates depend on assump­tions about how public opinion is likely to vary with demography and ge­ography, and there is no a priori right way to construct a multilevel mod­el of public opinion. One can always build a more complicated model, with more predictor variables and more interactions between pre­dictors. Consider again the (potential) relationship between racial atti­tudes and residential integration. As noted above, the analyst could ac­count for this by adding a measure of integration to the model for the intercept term. But what if the correlation between attitudes and residen­tial integration runs in opposite directions for highly educated and poorly educated whites? Highly educated whites who live in integrated neighborhoods are probably there by choice; poorly educated whites who end up in an inte­grated neighborhood may not be able to afford to move elsewhere and may feel threatened by the minority population. To ac­count for this pos­sibility, the researcher could construct a “varying slope” model, in which one of the demographic predictors (education) is inter­acted with an at­tribute of the geographic unit (residential integra­tion). But more elabo­rate models are not necessarily better. Researchers have shown that more complex MRP models sometimes yield worse esti­mates of target-population opinion, even though the complicated model does a better job explain­ing opinion within the pool of survey respond­ents. 261 Jeffrey R. Lax & Justin H. Philips, How Should We Estimate Sub-National Opinion Using MRP? Preliminary Findings and Recommendations 18–24 (Apr. 10, 2013) (unpublished manuscript), http://www.columbia.edu/~jrl2124/mrp2.pdf [http://perma.‌
cc/CK75-2ENZ] (finding tweaks to baseline model generate “modest gains at best and in some cases may actually reduce the performance of MRP”). This phenomenon, called overfitting, arises because the esti­mated pa­rameters in the more complex model capture idiosyncratic fea­tures of the sample that are not representative of the target population. 262 An important question for future work is whether machine-learning algorithms can be used to build and assess MRP models, automating this process rather than leaving it to the analyst’s discretion. If model-building is automated, this should allay concerns that the analyst calibrated the model to obtain results favorable to his or her client or political party.

There are, in principle, two ways of dealing with the overfitting prob­lem. One is to choose among candidate MRP models using a tech­nique known as cross-validation, whereby the data are randomly parti­tioned into, say, ten equally sized chunks; the model is fit ten times se­quentially using all but one of the data chunks; and the quality of the model’s pre­dictions are evaluated by comparing them to the actual ob­servations in the “left out” chunk of data. 263 For an introduction to cross-validation, see Trevor Hastie et al., The Elements of Statistical Learning: Data Mining, Inference, and Prediction 241–47 (2d ed. 2009). With sufficiently large sam­ples, cross-val­idation usually provides a fairly accurate estimate of true, out-of-sample prediction error—that is, how close the model’s predic­tions are likely to be to the actual observations in a new sample drawn at random from the population. 264 See id. at 247–49 (demonstrating cross-validation on hypothetical predictor and finding cross-validation error is close to true expected prediction error). But a recent paper raises questions about whether cross-validation reliably chooses the better MRP model. 265 Wei Wang & Andrew Gelman, Difficulty of Selecting Among Multilevel Models Using Predictive Accuracy, 7 Stat. & Its Interface 1, 1 (2014) (demonstrating while cross-validation might give satisfactory estimates of pointwise out-of-sample prediction error, it may not always be ideal for model comparison).

The other solution is to validate the MRP model with true, out-of-sam­ple observations of average public opinion (or a good proxy for aver­age opinion) within the geographic units and demographic strata of in­terest. 266 See supra note 260 and accompanying text (providing example of MRP predict­ing actual local referendum results). As noted above, MRP models of vote intention in presidential elections have been validated with data on the actual vote shares of the candidates in each geographic unit, and MRP models of public opinion on particular policy questions have been validated with vote-share data from initiative and referendum elections on similar policy proposals. 267 To be maximally convincing, however, the validation exercise should be done us­ing data obtained after the fitted model was placed into the public domain so that third parties can be confident that the validation data are truly “out of sample.” Otherwise, one cannot rule out the possibility that the researcher used some of the validation data to in­form his model-building choices, essentially reverse engineering the model to fit the vali­dation data. This standard is widely acknowledged in principle but rarely practiced. We are aware of no published work in political science that has validated an MRP model with data that were gathered after the model’s publication.

It is much trickier to validate MRP models concerning beliefs that are not voted on (such as general political ideology or racial stereotypes) or opinions within a group whose ballots are not separately tabulated (e.g., whites, Latinos, Asians, and blacks). One option is to assume that the model is reasonably good if it has a good theoretical justification and the same or similar models work well in predicting vote shares in the ge­ographic units. It seems unlikely that a model that performs well estimat­ing public opinion as a whole would do a bad job estimating within-group opinion, since the errors for each group would have to miracu­lously cancel out for the overall-opinion measure to be any good.

Relying on such assumptions is not ideal, but it is no more of a stretch than many other conventions of vote dilution litigation. Ecologi­cal inference as traditionally practiced relies on heroic assumptions, 268 A key assumption is that the political opinion of members of a racial group is un­correlated with the demographic makeup of the neighborhood in which the person lives. See generally Elmendorf et al., Racially Polarized Voting, supra note 60 (manuscript at pt. III.C, app. A) (“To estimate candidates’ vote shares by racial group, the analyst assumes that the proportion of white and minority voters who support each candidate is about the same in each precinct . . . .”). elides questions about statistical precision, and uses post-hoc corrections to paper over mathematically impossible results (such as an estimate that 130% of Latino voters supported candidate A over B). 269 See generally Greiner, Ecological Inference, supra note 78, at 162–63 (highlight­ing how ecological inference approach can result in logically impossible estimates). The analyst who uses an MRP model at least begins with individual-level data, and to the extent that she errs, she likely underestimates local deviation from typical patterns of opinion of the demographic group in question. 270 This is so because when local observations are sparse, the model pools the esti­mate of local opinion toward the typical opinion of the demographic type/unit type in the entire sample. For a lucid explanation and graphical illustrations, see Gelman & Hill, supra note 250, at 251–77. This seems the appropriate epistemic posture for a federal court implementing a federal statute: Assume that people in one part of the country are like people in another, except insofar as the data compel another conclusion.

The best way to validate an MRP model of within-racial-group opin­ion would be to conduct expensive, gold-standard surveys of public opin­ion within a randomly sampled subset of the geographic units. If the MRP predictions for each racial group in each unit are close to nonpar­ametric estimates from the validation study, the MRP model can be ad­judged “good.” 271 Short of this, one could make some headway on validation by estimating an at­tribute or behavior relevant to political opinion for which the true distribution within ra­cial groups and geographic units is known. The development of proprietary “big data” about voter registration and party preference by race presents one such opportunity. Cf. Stephen Ansolabehere & Eitan Hersh, Validation: What Big Data Reveal About Survey Misreporting and the Real Electorate, 20 Pol. Analysis 437, 445–58 (2012) (cross-referencing CCES self-reports of voter registration against proprietary information in Catalist database). Seen from one angle, the challenges of validation rep­resent serious obstacles to the proposed scheme of implementing section 2 with presumptions whose application to the case at hand would depend on MRP-generated estimates of racial group opinion. Seen from another, it represents a golden opportunity for DOJ.

Judicial coordination on datasets and models is key to getting a presumption-driven section 2 to do the work of section 5. If the courts agree on datasets and models, litigants (and judges) will be able to figure out quickly and easily which presumptions apply in a given case. DOJ has the time and resourcesR for gold-standard validation studies that could cost hundreds of thousands of dollars, or more. 272 UCLA political scientist Lynn Vavreck is in the process of developing new “gold-standard” methods for obtaining representative samples of electorate opinion. She pays respondents a lot of money, she offers part of the compensation as a gift, and she is trans­parent about the purpose of the survey. Lynn Vavreck, Presentation to Research Workshop in American Politics, University of California, Berkeley (Spring 2013). Most private litigants do not. If DOJ is the only player in the game whose model of racial group opinion has been validated with gold-standard surveys within a random sample of geographic units, judicial coordination on an MRP model is likely to occur much more quickly than if many actors (or no actors) have validated models. And it goes without saying that the “winning” model—the one on which courts eventually coordinate—is much more likely to be DOJ’s. Some defendants may still try to attack DOJ’s model by showing that its output is sensitive to modeling assumptions, but unless the defendant comes forth with better model that has been validated us­ing demonstrably out-of-sample data, 273 By “demonstrably out-of-sample,” we mean data collected after the fitted model was published. (This assumes that DOJ’s model passed an out-of-sample validation.) DOJ’s model is likely to prevail.

B. An Illustrative Model and Maps

To illustrate where a presumption-driven section 2 would probably have the most bite, this section reports original results on the geography of racial polarization in political preferences and racial stereotyping at the county level. The details of the model and replication code are avail­able online 274 Douglas M. Spencer, The Geography of Voter Discrimination, www.dougspencer.
org/research/geography_of_discrimination.html [http://perma.cc/N32H-HP‌8G] (last visited Oct. 21, 2015). and the online appendix also shows that the results are robust to alternative model specifications.

As section II.B.1 argued, the stereotyping results could be incorpo­rated into a “constitutional risk” presumption. The political-preference results could be used to establish whether minority preferences are suffi­ciently distinct from the preferences of others to support a vote dilution claim (potential disparate impact, as explained in section II.C.1). And, at least for judges who deem race discrimination for partisan purposes im­permissible, the preference–polarization results could also figure into constitutional-risk findings (section II.B.2).

The political preference measure used in this section is a one-dimensional ideal point scaled from respondents’ answers to binary policy questions on the 2010 Cooperative Congressional Election Study (CCES). 275 Respondents answered twenty-two policy questions. We convert their answers in­to thirty-eight actual or implied positions on binary choices. We generate ideal points us­ing the “ideal()” function of the “pscl” package in R. Simon Jackman, pscl: Political Science Computational Library, Stanford University, The Comprehensive R Archive Network (Mar. 29, 2015), https://cran.r-project.org/web/packages/pscl/index.html [http://perma.cc/SMK2-HH68]. For more information see Elmendorf & Spencer, Technical Appendix, supra note 37; see also Christopher S. Elmendorf & Douglas M. Spencer, An Intuitive Explanation of MRP, Administering Section 2 of the VRA After Shelby County, http://www.dougspencer.org/research/geography_of_discrimination/s2_‌Appendix_MRP.pdf [http://perma.cc/3CRA-AGN5] (last updated Jan. 7, 2015).
Recall from section II.C.1 that there are other plausible ways of measuring political pref­erences. Supra text accompanying notes 152–153. The results reported here are just illustrative. These ideal points are summary measures of liberalism or conservativeness as revealed by stated policy preferences. Ideal points calculated in this way have become standard fare in political science research on voter behav­ior, 276 See, e.g., Joseph Bafumi & Michael C. Herron, Leapfrog Representation and Extremism: A Study of American Voters and Their Members in Congress, 104 Am. Pol. Sci. Rev. 519, 521 (2010) (“Within studies of representation, there is movement toward com­paring voters’ preferences with legislator roll call voting behavior (i.e., with legislator ideal points).”); Cheryl Boudreau, Christopher S. Elmendorf & Scott A. MacKenzie, Lost in Space? Shortcuts and Spatial Voting in Low-Information Elections, 68 Pol. Res. Q. (forthcoming 2015) (manuscript at 10), http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2232371 (on file with the Columbia Law Review) (incorporating estimation of ideal points for candidates and voters in study design); Stephen A. Jessee, Partisan Bias, Political Information and Spatial Voting in the 2008 Presidential Election, 72 J. Pol. 327, 328 (2010) (discussing voters’ ideal points and spatial utility model of voting); Stephen A. Jessee, Spatial Voting in the 2004 Presidential Election, 103 Am. Pol. Sci. Rev. 59, 67 (2009) (noting bill parameters in ideal point model are identical for respondent answer­ing survey question and for senator voting on pro­posal); Tausanovitch & Warshaw, supra note 246, at 231–32 (“We assume that both citizens and legislators have a unique set of policies that they ‘prefer’ to all others. This point in the policy space is called an ‘ideal point.’”); Christopher Hare & Keith T. Poole, Using Optimal Classification to Analyze Public Opinion Data 3–12 (Mar. 16, 2012) (unpublished manuscript) (on file with the Columbia Law Review) (explaining use of scaling methods to estimate voters’ ideal points); Thad Kousser, Justin Phillips & Boris Shor, Reform and Representation: A New Method Applied to Recent Electoral Changes 8 (Aug. 19, 2015) (unpublished manuscript), http://ssrn.com/abstract=2260083 (on file with the Columbia Law Review) (describing methodology of “measuring . . . ideology of a sample of voters . . . by comparing their an­swers to the same comprehensive set of policy questions” to obtain “common space ideo­logical estimates”); Boris Shor, Congruence, Responsiveness, and Representation in American State Legislatures 4 (Aug. 25, 2014) (unpublished manuscript), http://ssrn.com/
abstract=1697352 (on file with the Columbia Law Review) (describing use of “standard ideal point estimation techniques” to construct “common ideological space” for politicians and constituents); Boris Shor & Jon C. Rogowski, Ideology and the US Congressional Vote 12, 13 & tbl.2 (July 15, 2015) (unpublished manuscript), http://‌ssrn.com/abstract=2650028 (on file with the Columbia Law Review) (detailing surveys used to “characterize the ideo­logical locations” of survey respondents). and they explain much more of the variation in vote choice than does self-reported ideology. 277 See Shor & Rogowski, supra note 276, at 12–13 (explaining ideal points method avoids problem of “projection” because respondents are “unlikely to adopt the issue posi­tion of their favored local House candidate” on significant policy issues). Ideal points scaled from policy positions are particularly valuable for comparing the political preferences of racial groups, because there is considerable between-group variation in how re­spondents characterize their own ideology on the liberal-to-conservative spectrum 278 See Marisa Abrajano, Reexamining the “Racial Gap” in Political Knowledge, 77 J. Pol. 44, 46 (2015) (showing many Latino American survey respondents interpret word “liberal” to mean “conservative,” reflecting different use of term “liberal” in their country of origin or ancestry). In the technical online appendix to this Article, we demonstrate that the “racial gap” in the correlation between respondent ideology and preferences in con­gressional and presidential elections is smaller when ideology is measured using ideal points than when ideology is measured using self-reports on the seven-point scale. Elmendorf & Spencer, Technical Appendix, supra note 37, at 9–10 & tbl.1. and in their willingness to express a party identification. 279 See Zoltan L. Hajnal & Taeku Lee, Why Americans Don’t Join the Party: Race, Immigration, and the Failure (of Political Parties) to Engage the Electorate 89 tbl.3.2, 108–10 tbls.4.2 & 4.3 (2011) (showing Asian American and Hispanic American survey re­spondents are much less willing to express party identification than African American respondents). The scaled-ideal-point measure of political pref­erences is imperfect, 280 It is imperfect (1) because it does not account for the importance that respond­ents attach to different issues; (2) because some of the variation in scaled ideal points prob­ably reflects differences in political knowledge rather than differences in latent ideol­ogy (low-knowledge respondents may make more “errors” in stating their policy positions, see Thomas R. Palfrey & Keith T. Poole, The Relationship Between Information, Ideology, and Voting Behavior, 31 Am. J. Pol. Sci. 511, 529–30 (1987) (concluding low-information voters are more likely to be indifferent and unpredictable in voting behavior)); (3) be­cause respondents who have ideal points “in the middle” may not agree with one another very much, see Broockman, supra note 170, at 29–30 (observing individuals appearing in ideological middle sometimes have “dramatic differences” in preferences with respect to specific policy domains); (4) because ideal points scaled with a parametric model (as is conventional, and as we do here) may be sensitive to the set of policy questions used in the analysis and to functional-form assumptions about voters’ utility functions, see Hare & Poole, supra note 276, at 6–7 (noting parametric models impose constraint “that mass po­litical attitudes are uniformly structured across the electorate[,]” entailing assumptions about voters’ utility curves); (5) because ideal points scaled from national political issues may not capture preferences over local politics, and as such may be poorly suited to VRA claims concerning school district or city council elections, etc., see Boudreau et al., supra note 276 (manuscript at 21–22) (showing relatively weak correlation between voters’ national party identification and their ideal points in issue space of San Francisco politics); David Schleicher, Why Is There No Partisan Competition in City Council Elections?: The Role of Election Law, 23 J.L. & Pol. 419, 433–44 (2007) (arguing this is likely); and (6) because ideal points do not capture variation in political preferences that arise from and reflect distributional politics. (Thanks to David Schleicher for pointing out this final limitation.) but it beats self-reported ideology and party identification.

To measure ideological similarity within and between racial groups, we use the average ideological distance (absolute value) between pairs of citizens of the group or groups in question, divided by the average ideo­logical distance between pairs of citizens chosen at random from the en­tire population. 281 In mathematical notation: ave|x_Ai-x_Bj |⁄ ave|x_k-x_l | i∈A,j∈B and . In this for­mula ave is the average (mean) operator, x is an ideal point, A and B index racial groups (the populations of which are A and B ), and P  is the entire population. When measuring within group cohesion, A =B and i ≠j . A score greater than one indicates polarization, as it signifies that the typical distance between two members of the group(s) in question is greater than the typical distance between a pair of citizens in the population as a whole. Conversely, a score less than one indicates cohesion, that is, greater similarity within the numerator group(s) than in the full population. 282 Because ideal points establish relative but not absolute distances between voters, it is necessary to standardize the “similarity” measure in some way (such as by dividing by the standard deviation of the distance between randomly sampled pairs of voters in the entire national population).

For VRA purposes, one very nice feature of this approach is that it can be used to answer (presumptively) all of the group cohesion ques­tions in a vote dilution case. “Minority political cohesion” is assessed by sampling same-race pairs of voters for the numerator; “polarization” is assessed by sampling different-race pairs. Coalitional claims brought jointly by two or more racial groups (which have vexed the courts) pre­sent no special difficulty. 283 As the United States becomes more racially diverse and residentially integrated, co­alitional claims will become increasingly important for minority representation, because few racial communities will be able to satisfy the “majority-minority” requirement of Gingles—as glossed by Bartlett v. Strickland, 556 U.S. 1 (2009) (plurality opinion) (Kennedy, J.)—on their own. Two racial groups are presumptively jointly cohesive if the typical distance between randomly selected pairs of voters from the two groups falls below whatever threshold the courts may estab­lish for political cohesion in ordinary, noncoalitional cases. (To be sure, it is an open question whether coalitional claims remain available under section 2 after Bartlett v. Strickland. 284 Bartlett forecloses “crossover” claims brought by minorities who do not compose 50% of the proposed remedial district but would be able to control the district together with white allies. Id. at 14–15 (plurality opinion) (Kennedy, J.) (“Nothing in § 2 grants special protection to a minority group’s right to form political coalitions.”). Though Bartlett expressly reserves the question of coalitional claims brought by two minority groups who together surmount the 50% threshold, some commentators not unreasonably see Bartlett as foretelling the demise of coalitional claims. See, e.g., Lauren R. Weinberg, Note, Reading the Tea Leaves: The Supreme Court and the Future of Coalitional Claims Under Section 2 of the Voting Rights Act, 91 Wash. U. L. Rev. 411, 433 (2013) (concluding Bartlett and Perry v. Perez, 132 S. Ct. 934 (2011), “strongly suggest that if and when the issue of coalition districts is directly pre­sented,” Court will hold they are not “afforded protection under section 2”). On the other hand, coalitional claims present none of the line-drawing problems that so troubled the Bartlett plurality, nor do they depend on the existence of serious fractures in the white voting “bloc” (fractures which, as Bartlett ob­serves, would create “serious tension” with the third prong of Gingles, see Bartlett, 556 U.S. at 16 (plurality opinion) (Kennedy, J.)). Either the two minority groups together compose 50% of the proposed remedial district or they do not. Either they are jointly po­litically cohesive or they are not. This is quite different than the “crossover”-type claim at issue in Bartlett, in which a minority group that composes a minority of the proposed re­medial district claims the ability to control it in coalition with some indeterminate sub-population of the less-than-fully-cohesive white majority. The lower courts have split on the availability of coalitional claims, with most courts allowing them in principle, but the rele­vant deci­sions mostly predate Bartlett. See Hopkins, supra note 35, at 635–36 (summarizing cases). )

A key conceptual question is how to define “the entire population” for purposes of calculating the denominator of the cohesion/polarization measure. Should the similarity/dissimilarity of political preferences in the numerator group(s) be measured relative to typical similarity within the population that elects the legislative body at issue in the case or rela­tive to the entire national population? We think the former approach probably makes more sense. If the citizens of, say, a county, divide politi­cally on racial lines alone, the minority community may have a very hard time electing the county commissioners it prefers even if the typical ideo­logical distance between minority and majority-race voters in the county is smaller than the typical distance between any two voters in the national population. 285 This assumes that voters figure out the ideological positions of candidates in the county commissioner elections and that citizens’ ideological positions in national and local politics are highly correlated. Both assumptions are questionable. See generally Christopher S. Elmendorf & David Schleicher, Informing Consent: Voter Ignorance, Political Parties, and Election Law, 2013 U. Ill. L. Rev. 363 (discussing lack of voter knowledge about candi­dates’ positions, particularly in subnational elections); Schleicher, supra note 280 (arguing voter ideology in local and national issue spaces is likely to be only weakly correlated). But for present purposes, it is enough to provide a simple, easily interpreted picture of geographic variation in racial polarization throughout the nation, so the “national population” de­nominator will be used. 286 We use the Integrated Public Use Census Microdata as our estimate of the na­tional population. Steven Ruggles, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder & Matthew Sobek, Integrated Public Use Microdata Series: Version 5.0, Univ. of Minn. (2010), https://usa.ipums.org/usa/index.shtml [https://perma.cc/‌Y65N-KRWJ].

One other complication needs to be mentioned. The model esti­mates the mean ideal point for “types” of voters defined by the poststrati­fication cells (race, age, sex, education, and geographic unit). It does not give the full distribution of ideal points within small geographic units, which can be obtained only by conducting large surveys within each unit. However, by sampling pairs of voters from the poststratification cells, and imputing to them the mean ideal point estimated for the cell, one can still generate a picture of the relative distance within and between voters of different groups. 287 In this exercise, the probability of choosing a voter in a given cell equals the rela­tive frequency of voters in that cell. This will tend to understate the actual diversity of opinion within the population (because sampling is done from subgroup means, rather than individuals), but so long as mean ideal points from the poststratification cells are used for the denominator, the model will capture whether between-racial-group differences are larger or smaller than differences across the full set of demographic cleavages in the MRP model (age, sex, race, education, and geography). 288 In future work, we will pursue another strategy that may better recover the diver­sity of public opinion: modeling the ideological distance between pairs of voters of differ­ent types, rather than the mean ideal point of each voter type. Still another possibility is to sample from the residuals of an estimated model of mean opinion and use bootstrap meth­ods to estimate between-group opinion.

To estimate the mean ideal point of each voter type, we first subset voters by race and fit separate models for each racial group. This allows coefficients on the predictor variables to vary across racial groups, with­out any “pooling” of information between groups. Put differently, we do not try to model potential commonalities across racial groups (e.g., by positing that the correlation between income and ideology is similar for each racial group), and we do not use ideal points of persons of race A to predict ideal points for persons of race B. Our county-level models in­clude age and sex as individual-level predictors and state as an aggregate predictor. We also include one county-level attribute: the minority per­centage of the county’s population. 289 This is motivated by the “racial threat hypothesis,” which posits that members of the majority group subscribe to worse views of the minority where the minority threatens the privilege or advantages of the majority. See supra note 252 (discussing racial threat hypothesis). For a review of this literature and some interesting new results, see Ryan D. Enos, What the Demolition of Public Housing Teaches Us About the Impact of Racial Threat on Political Behavior, 109 Am. J. Pol. Sci. (forthcoming 2015) (manuscript at 1–2, 11–16) (on file with the Columbia Law Review) (studying impact of exogenous removal of black community caused by reconstruction of Chicago public housing on white voting, finding white turnout fell over 10% and “pro-Republican leanings of voters near projects” declined after projects demolished, supporting racial threat hypothesis).

Figure 1 maps the results on ideological polarization between white and minority citizens at the county level. It shows that the ideological gap between white and black citizens in most counties is vastly greater than the gap between whites and Asians, and whites and Latinos. The white–black gap is most pronounced in the South, where the typical distance between whites and blacks is often twice as large as the typical distance between voting-age Americans as a whole. There is also significant white–Asian and white–Latino polarization in Texas and in a scattering of coun­ties elsewhere, mostly but not entirely in the South.

Figure 1: County-Level Estimates of the Divergence Between the Average White Voting-Age Citizen’s Ideal Point and the Average Ideal Points of Black, Latino, and Asian American Voting-Age Citizens 290 Data: 2010 Cooperative Congressional Election Study (N=47,234).
Elmendorf Maps 1

Figure 2 uses the same ideal point data and MRP models to illus­trate similarities between minority groups. 291 In the technical online appendix, we replicate the method of Figures 1 and 2 to map the geography of within-group political cohesion. The results are not particularly interesting: within counties, the estimated typical distance between voters of a given racial group is uniformly less than one half of the typical distance between voting-age Americans as a whole. Elmendorf & Spencer, Technical Appendix, supra note 37, at 12, 13 & tbl.2. Figure 2 drives home that coalitional claims brought by two or more minority groups can be ana­lyzed in essentially the same way as claims brought by a single racial group. The courts, aided by DOJ, just need to set a quantitative threshold for what constitutes “substantial similarity” using the measure of cohe­sion/polarization. A cutoff of 0.5, for example, would require voters of each minority group to be twice as close to one other, on average, as vot­ers in the full population. Minority groups that fall on the “similar” side of the threshold would be deemed presumptively jointly cohesive. 292 Those on the far side could not bring a coalitional claim unless they introduce new survey or voting data to overcome the presumption of noncohesiveness.

Figure 2: Difference in Ideal Points Between Minority Groups 293 Data: 2010 Cooperative Congressional Election Study (N=47,234).
Elmendorf Maps 2

The main takeaway from Figure 2 is that the two fastest growing ra­cial groups in the United States, Asian Americans and Latinos, are ideo­logically very similar. 294 Note that our models do not distinguish among Latino populations by national origin. We pool information across Cuban Americans, Mexican Americans, Puerto Ricans, and others. One may fairly question whether the resulting estimates are reliable for Cuban Americans, who may be more conservative on average than the other Latino populations. Cf. De Grandy v. Wetherell, 815 F. Supp. 1550, 1570–72 (1992), aff’d in part, rev’d in part sub nom. Johnson v. De Grandy, 512 U.S. 997 (1994), in which the trial court found that African Americans and Cuban Americans in Dade County, Florida, had very different po­litical preferences. In almost every county, the distance between Asian American and Latino voters is less than half the typical distance between voting age Americans as a whole. To date, however, Asian–Latino coalitional claims under section 2 have been rare—and rarely suc­cessful. 295 See Chen & Lee, supra note 35, at 391 (describing “paucity of Asian American-focused litigation”). If the presumption-based approach to section 2 were adopted, many of these claims might become winners. 296 At least if the constitutional risk requirement can be satisfied. Given space limita­tions, the question of how this requirement should be understood in coalitional cases can­not be addressed here.

The map in Figure 3 summarizes racial stereotyping against blacks, Latinos, and Asians at the county level. 297 Formally, county-level stereotyping against group M is: “\”(ave(S_iK^M))/(ave(S_j))| i∈K,j∈P,{i,j}∉M\” ” , where S is an individual’s stereotype, i and j index voting-age citizens, S  is the voting-age citizen population of the county of interest (K), M is the racial group being stereotyped, M  is the population of that racial group, P  is the national population of voting-age citizens, and the averages are taken over K  and P  (exclusive of M ), respectively. In turn, S_m=sum(R_izm – R_iz) over i where R_izm is respondent i’s rating of minority group M on attribute z (e.g., intelligence), and R_iz0 is the respondent’s rating of his or her own racial group on the same attribute. The map is color coded to re­flect the proportion of residents in each county that harbor a substan­tially negative stereotype of each racial group. We distinguish substantial from insubstantial negative stereotyping based on the predictive value of racial stereotypes with respect to vote choice. More specifically, we model the probability of voting for Obama in 2008 as a function of anti-black stereotyping, political ideology, party identification, age, sex, race, educa­tion, income, and region among respondents to the 2008 Annenberg National Election Survey and the 2008 Cooperative Campaign Analysis Project survey. 298 The balance of this paragraph summarizes the modeling and results in Elmendorf & Spencer, Preclearance, supra note 19, at 1143–56. Respondents who view blacks more positively than whites were no more or less likely to vote for Obama than the national median voter. Respondents whose views of blacks are more than one standard deviation more negative than the national average were much less likely to vote for Obama than Hillary Clinton in the 2008 primary, less likely to vote for Obama than John McCain in the 2008 general election, and much more likely to be “Democratic defectors” who voted for John Kerry in 2004 and McCain in 2008. Based on this relationship, we use the anti-black stereotyping value that is one standard deviation more negative than the national average (a value of 2.49 on a 12-point scale) as the cut­off for identifying substantial negative stereotyping. Figure 3 plots the proportion of respondents in each county that exceed this cutoff.

Figure 3: County-Level Differences in the Proportion of Non-Coethnic Survey Respondents with Substantially Negative Stereotypes of Each Minority Group 299 Data: 2010 Cooperative Campaign Analysis Project Survey (N=19,187).
Elmendorf Maps 3

 

There are striking geographic patterns to the stereotyping of blacks, Asian Americans and Latinos. Non-blacks in the former Confederacy har­bor the most negative stereotypes about blacks. Anti-Latino stereotyp­ing is strongest in New Mexico, Arizona, Southern California, and Southwestern Texas. Substantial anti-Asian stereotyping is much less prev­alent across the country, yet relatively strong in New Mexico and a few counties in the upper Midwest.

Read together, Figures 1, 2, and 3 contain important lessons about what might emerge as the de facto “coverage formula” of a presumption-driven section 2—that is, the geographic regions in which the presump­tions would operate to shift core evidentiary burdens to the defendant. For purposes of claims brought by African Americans, the presumptions’ coverage would be substantially similar to section 5’s coverage pre-Shelby County. Black–white ideological polarization and negative stereotyping of blacks are both concentrated in the Deep South. For claims brought by Latinos and Asians, the picture is more complicated because negative stereotyping and ideological polarization with respect to these groups apparently do not go hand in hand. Asians face the worst stereotypes in the upper Midwest, but are most polarized (vis-à-vis whites) in Texas. Anti-Latino stereotyping and ideological polarization both occur in Texas and to some extent in the upper South, but are not geographically con­cordant elsewhere. So even though Asians and Latinos represent good “coa­litional partners” under section 2 by dint of their ideological similar­ity, they may have difficulty satisfying the constitutional-risk ele­ment of a section 2 claim. 300 Even if the courts accept the “partisan incentives” arguments about constitu­tional risk, the results indicate that white–Latino and white–Asian ideological polarization (and hence political incentives to discriminate) rarely reaches the levels of white–black ideological polarization. Supra Fig.1.

C. Next Steps

The modeling and results reported here speak to the feasibility of im­plementing section 2 with rebuttable presumptions whose application in a given case would be determined using national survey data, but this Article does not establish the optimality of any particular presumption, measure of preferences, or predictive model. There is a good deal of re­search still to be done on these questions and this Part closes by enumer­ating the most important empirical projects for better grounding and targeting a presumption-driven section 2:

  • Develop other plausible measures of racial-group political prefer­ences and assess the sensitivity of geography-of-polarization/cohesion results to the choice of measure. 301 We touch briefly on several alternative measures in section II.C.1, supra notes 152–153. Regarding the strengths and limitations of the ideal point measure used in this Article, see supra notes 276–280 and accompanying text.
  • Investigate whether racial-group political preferences can be well represented with a single summary measure that is independent of the governmental body at issue or whether different classes of governments (e.g., school boards, city councils, state legislatures, Congress) require different measures. 302 See supra note 280 and accompanying text (noting potential imperfections of measuring political preference by ideal points).
  • Investigate sensitivity of geography-of-polarization results to alter­na­tive specifications of the MRP model 303 Cf. supra note 288 (discussing alternative approach to estimating ideological po­larization within small geographic units). and explore machine-learning algorithms for “impartial” model selection and specification. 304 See supra note 262. Machine-learning methods can make model specification less dependent on the analyst’s judgment calls. See generally Bertrand Clarke, Ernest Fokoue & Hao Helen Zhang, Principles and Theory for Data Mining and Machine Learning (2009) (explaining machine-learning methods and model building). As such, they can help to allay concerns that a particular model was chosen because the analyst “liked” its results.
  • Replicate geography-of-discrimination results using alternative measures of racial attitudes. 305 It would be particularly helpful to have a behavioral measure of disparate treat­ment or differential sympathy, one that is less vulnerable to social-desirability biases than the explicit stereotyping measure used here, and that tracks the equal protection norms against disparate treatment. For an explanation of why we rely on explicit stereotyping measures for the time being, see Elmendorf & Spencer, Preclearance, supra note 19, at 1143–44. Historical and cultural studies also suggest that Asian Americans are stereotyped differently than blacks and Latinos. Specifically, Asians are often represented as “perpet­ual foreigners” and “model minorities.” See Gary Okihiro, Margins and Mainstreams 134–42 (1994) (describing historical characterization of Asian Americans as “yellow peril” and “model minority”); Claire Jean Kim, The Racial Triangulation of Asian Americans, 27 Pol. & Soc’y 105, 106, 118 (1999) (noting triangulation of Asian Americans as “outsiders or aliens” and as “model minority” compared to blacks). We are grateful to Fred Lee for directing us to this literature. Future studies should assess the sensitivity of our geography-of-discrimination results to these alternative measures of attitudes towards Asians.
  • Use MRP or other techniques to estimate the frequency of regis­tered voters and likely voters within the poststratification cells of an MRP model of racial group opinion. (This would make it pos­sible to ground section 2 presumptions in the distribution of opin­ion among registered or likely voters, as opposed to the citi­zen voting-age population. 306 Poststratification by the population of registered or likely voters requires this prior modeling exercise because the Census Bureau asks only a small sample of Americans about voting and registration. The Voting and Registration Supplement is included in the November Current Population Survey, a survey of approximately 50,000 households. Voting and Registration, U.S. Census Bureau, http://www.census.gov/hhes/www/soc‌demo/voting/ [http://perma.cc/W5EU-XELG] (last modified July 16, 2015). )
  • If resources are available, validate MRP models with large-N sur­veys in a randomly selected subset of jurisdiction, 307 See supra notes 235–267 and accompanying text (suggesting techniques for esti­mating opinions of racial groups in specific geographic units). or (second best) with studies of a politically relevant behavior or attitude with respect to which the joint distribution of race and the behav­ior/attitude in the population can be treated as known. 308 See supra note 271 (suggesting data about voter registration and party prefer­ence by race could be one such attribute).
  • Develop protocols for converting MRP-generated presumptions into informative priors for use with Bayesian ecological inference models. 309 This is one way to structure litigants’ attempts to overcome the presumptions. Bayesian methods have been used in most recent work by leading political methodologists on ecological inference. E.g., Glynn & Wakefield, supra note 80; Greiner & Quinn, R×C Ecological Inference, supra note 219; Ori Rosen et al., Bayesian and Frequentist Inference for Ecological Inference: The R×C Case, 55 Statistica Neerlandica 134 (2001). But we are aware of no work to date on the establishment of informative priors for use in Bayesian Ecological Inference models.
  • Develop models to estimate minority opportunity in non-major­ity-minority districts, so that presumptions about which districts are minority-opportunity districts can be refined. 310 Cf. Charles S. Bullock, III & Richard E. Dunn, The Demise of Racial Districting and the Future of Black Representation, 48 Emory L.J. 1209, 1240–41 (1999) (estimating white crossover voting for black candidates in congressional races); Bernard Grofman et al., Drawing Effective Minority Districts: A Conceptual Framework and Some Empirical Evidence, 79 N.C. L. Rev. 1383, 1407–09 (2001) (same).

While these research projects could greatly strengthen a presump­tion-driven section 2, they are not essential to get it up and running. Vote dilution cases have long been litigated using somewhat shaky methods of ecological inference and the courts have muddled through. 311 See supra notes 218–220 and accompanying text (considering Greiner’s cri­tiques of ecological inference and lack of attention paid by courts). The imple­mentation of a presumption-driven section 2 can proceed similarly.

Conclusion

Numerous commentators in the wake of Shelby County offered pro­posals for putting section 5 back to work with new coverage formulas or “bail-in” remedies. 312 See, e.g., Elmendorf & Spencer, Preclearance, supra note 19, at 1127–28 (pro­posing new section 5 coverage formula); Bernard Grofman, Devising a Sensible Trigger for Section 5 of the Voting Rights Act, 12 Election L.J. 332, 334 (2013) (suggesting new “trig­ger” conditions for section 5); Morgan Kousser, Gutting the Landmark Civil Rights Legislation, Reuters: Analysis & Opinion (June 26, 2013), http://blogs.reuters.com/great-debate/2013/06/26/gutting-the-landmark-civil-rights-legislation [http://perma.cc/36X‌X-XJ9U] (recommending updated coverage formula and automatic “bail in” solutions); Spencer Overton, How to Update the Voting Rights Act, Huffington Post (June 25, 2013, 12:52 pm), http://www.huffingtonpost.com/spencer-overton/how-to-update-the-voting_b_3497350.html [http://perma.cc/W23X-L97T] (“Congress should update the coverage for­mula to require that states and localities with recent voting rights violations preclear new elec­tion law changes.”); Richard Pildes, One Easy, but Powerful, Way to Amend the VRA, Election Law Blog (June 28, 2013, 6:53 am), http://‌electionlawblog.org/?p=52349 [http://perma.
cc/K7M3-NWGU] (advocating revi­sion of section 3’s “bail-in” provision). This Article suggests another tack: Establish eviden­tiary presumptions under section 2 that shift the burden of proof to de­fendants in political jurisdictions where the risk of unconstitutional race discrimination with respect to voting is elevated. Where the presump­tions apply, section 2 would function much like section 5: It would be fairly easy to block potentially discriminatory electoral reforms before they take effect and lawmakers would have correspondingly strong ex-ante incentives to gather information about the likely effects of reforms on racial minorities and to mitigate adverse impacts.

The empirical results reported in this Article suggest that a presumption-driven section 2 would “cover” most of the Deep South for purposes of claims brought by African Americans, much like section 5 prior to Shelby County. But the picture is more complicated for Asian Americans and Latinos. These groups, while jointly politically cohesive, are generally less ideologically polarized vis-à-vis whites than are African Americans, and the areas of greatest white–Asian and white–Latino polarization are not (consistently) the areas with the most negative stereotyping of Asians and Latinos.

Perhaps the greatest challenge for realizing a presumption-driven section 2, implemented with national survey data, is getting the courts to define and operationalize the presumptions similarly. Congress could solve this problem by authorizing DOJ to issue substantive rules under section 2. Even without a grant of rulemaking authority, DOJ may be able to induce judicial coordination by issuing guidance documents, which would be owed Skidmore deference, and by sponsoring model-validation studies. And if DOJ stays on the sidelines or is rebuffed by the courts, case-by-case litigation using national survey data has some potential to gradually reform the law of section 2, in ways that reduce the cost and increase the predictability of voting rights enforcement.